Renewable and Sustainable Energy Reviews 198 (2024) 114391

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

journal homepage: www.elsevier.com/locate/rser

Review article

A review of distributed solar forecasting with remote sensing and deep

learning
Yinghao Chu a , Yiling Wang b,c , Dazhi Yang d , Shanlin Chen b , Mengying Li b ,∗
a
Department of Systems Engineering, City University of Hong Kong, Hong Kong Special Administrative Region
b
Department of Mechanical Engineering & Research Center on Data Science and Artificial Intelligence, The Hong Kong Polytechnic University, Hong Kong Special
Administrative Region
c
School of Energy and Environment, City University of Hong Kong, Hong Kong Special Administrative Region
d
School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin, Heilongjiang, China

ARTICLE INFO ABSTRACT

Keywords: The rapidly growing capacity of globally distributed solar generation systems (DSGs) has imposed new
Review challenges for solar forecasting research: the need for high-fidelity spatial solar forecasts across utility-
Solar integration scale areas with minimized capital, generalization, and maintenance costs. The majority of solar forecasting
Spatial solar forecasting
approaches were developed for centralized solar power plants, which only concern one or a few locations.
Remote sensing
Therefore, this work reviews the state-of-the-art methods for spatial solar forecasting that integrate deep
Deep learning
Hybrid methods
learning and remote sensing, potentially capable of serving numerous DSGs simultaneously. This work has four
missions: (1) provide a review of available remote-sensing- and deep-learning-based spatial solar forecasting
methods; (2) provide suggestions of practical tools to accelerate the research and deployment of spatial solar
forecasting methods; (3) identify challenges of spatial solar forecasting for sparsely distributed DSGs; and (4)
discuss prospective approaches to further enhance both the performance and value of spatial solar forecasts,
such as the attention mechanism, sequence analysis, or probabilistic forecasts. This work reveals that practical
spatial solar forecasting for DSGs is still in its infancy, thus more research efforts should be involved to develop
a new generation of forecasting engines, which could cost-effectively address the real-time needs of integrating
massive regional DSGs.

1. Introduction investment entry barriers, local grid stabilization, and flexible installa-
tion options that optimize land use, such as rooftop or floating PV [7,8].
Motivated by the societal need for sustainable development around Therefore, recent installations of DSGs have markedly outpaced those of
the world, the global installed capacity of solar photovoltaic (PV) centralized solar systems, especially in densely populated regions with
has grown rapidly over the past decade and is expected to increase limited land availability [9]. With hundreds of grid-connected DSGs
by 20-fold and reach 22% of the electricity market by 2050 [1,2]. in operation [10], regional-scale spatial solar forecasting is essential
A variety of studies have confirmed that the accurate forecasting of for the planning, integrating, regulating, and managing of solar power
weather-dependent solar fluctuations is key to integrating the volatile generations over a large area [11].
and non-dispatchable solar energy into the power grid [3–5]. How- Motivated by the increasing demands of solar integration, different
ever, accurately forecasting either solar irradiance or solar power is
solar forecasting approaches have been proposed and discussed in
a challenging task because of the complex meteorological processes.
the literature, such as numerical weather prediction (NWP), physical
Specifically, the properties and dynamics of clouds, which exhibit high
models that are based on either local- or remote-sensing techniques,
spatial and temporal variability, contribute to this complexity [6].
data-driven methods based on regressive, machine learning, and deep
In addition, the rapidly growing installation of PV systems, particu-
larly distributed solar generation systems (DSGs) over large areas, has learning approaches, and hybrid approaches that integrate multiple
imposed new challenges to solar forecasting and grid integration. Com- methods to maximize the forecasting performance [2,3,12]. The com-
pared to centralized solar energy systems, DSGs offer several benefits monly employed data sources include in-situ measurement of solar
including proximity to consumers, minimized transmission loss, lower irradiance [13] and DSGs power generation [14], sky images [11,15],

∗ Corresponding author.
E-mail address: [email protected] (M. Li).

https://doi.org/10.1016/j.rser.2024.114391
Received 24 November 2022; Received in revised form 13 November 2023; Accepted 24 March 2024
Available online 25 April 2024
1364-0321/© 2024 Elsevier Ltd. All rights reserved.
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Nomenclature RNN Recurrent neural network

ROI Region of interest
Abbreviations rRMSE Relative root mean square error
SARIMAX Seasonal autoregressive integrated moving average
ABI Advanced Baseline Imager
with exogenous factors
ADAM Adaptive moment estimation
SEVIRI Spinning Enhanced Visible and infrared Imager
AGRI Advanced Geostationary Radiation Imager
SVM Support vector machine
AHI Advanced Himawari Imager
SVR Support vector regression
AMI Advanced Meteorological Imager
WRF Weather Research and Forecasting
ANN Artificial neural network
XGBoost Extreme gradient boosting
ARIMA Autoregressive integrated moving average
biLSTM Bidirectional long short-term memory Notations
CCI Cloud clearness index
𝛼 Coefficient of relative positions between satellite
CMA China Meteorological Administration and target location
CNN Convolutional neural network 𝛾ℎ The autocorrelation at lag ℎ
COMS Communication, Ocean and Meteorological Satellite 𝜅̂ Clear-sky index forecast
DES Double exponential smoothing 𝐼̂ Solar irradiance/power forecast
DHI Diffuse horizontal irradiance 𝜅 Clear-sky index
DNI Direct normal irradiance 𝜇 Climatology reference term
DNN Deep neural network 𝜌 Correlation coefficient
DSG Distributed solar generation 𝜎 Standard deviation
ECMWF European Centre for Medium-range Weather Fore- 𝜃𝑧 Solar zenith angle
casts
𝐿̃ Normalized pixel values
ELM Extreme learning machine
𝑐 Clear-sky expectation of irradiance/power
EMOS Ensemble model output statistics
𝐻𝑎 Rényi entropy
EUMETSAT European Organization for the Exploitation of
𝐼 Solar irradiance/power
Meteorological Satellite
𝐿 Pixel value of satellite images
FFNN Feed-forward neural networks
𝐿′ Normalized pixel values by solar zenith angle
GAN Generative adversarial network
𝑁 Number of days
GBR Gradient boosting regression
𝑅2 Coefficient of determination
GFS Global Forecast System
𝑠 Forecast skill
GHI Global horizontal irradiance
GK2A/2B Geostationary Korean Multi-purpose Satellite 2A/2B Subscripts
GNN Graph neural network 𝑓 Forecasts of interest
GOES Geostationary Operational Environmental Satellites ℎ Forecast horizon
GRAPES Global and Regional Assimilation and Prediction 𝑖 Number of row in a matrix
System
𝑗 Number of column in a matrix
GRU Gated recurrent unit
𝑝 Perfect forecasts
HARMONIE Hirlam–Aladin Research towards Mesoscale Opera-
𝑟 Reference forecasts
tional NWP In Euromed
𝑡 Time
HRRR High-Resolution Rapid Refresh
KMA Korea Meteorological Administration Superscripts
kNN 𝑘 Nearest neighbor
𝑏 Spectral band
KSI Kolmogorov–Smirnov integral
LSTM Long short-term memory
LSTM-FC Long short-term memory-fully connected
MAE Mean absolute error
MBE Mean bias error meteorological variables (e.g., temperature, relative humidity, wind
MLP Multilayer perceptron direction or wind speed) that are measured or predicted by NWP
MODIS Moderate resolution imaging spectroradiometer models [16–18], and remote-sensing satellite data products [19–21].
NAM North American Mesoscale For solar forecasting research, the spatial and temporal scales of the
aforementioned data sources are illustrated in Fig. 1. For spatial fore-
NMSC National Meteorological Satellite Center
casts over hundreds of square kilometers, the choices of data are
NSRDB National Solar Radiation Database
remote-sensing imagery and NWP output. Since comprehensive details
NWP Numerical weather prediction of state-of-the-art NWP models are widely accessible elsewhere, such
PV Photovoltaic as the articles by Wang et al. [22], Yang et al. [23] and Jimenez et al.
RAP Rapid Refresh model [24], this work only briefly summarizes a few NWP models commonly
ReLU Rectified linear unit used in solar forecasting research, as presented in Appendix A.
RF Random forest Remote sensing has emerged as a reliable and informative data
rMAE Relative mean absolute error source for solar resource assessment and forecasting. Irradiance derived
RMSE Root mean square error thereof has wide geographical coverages and thus partly eliminates
the need for local instrumentation investment [26]. Since the presence

2
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

year 2022, searching keywords such as deep learning with solar forecast
or PV forecast returns more than 2000 publications from the Web
of Science Core Collection. However, when searching deep learning
with spatial solar forecast, the number of return publications decreases
to just 45, among which a fraction use satellite data. Consequently,
how to best perform satellite-based spatial solar forecasting using deep
learning techniques has not been standardized and the related research
gaps have not been identified.
Therefore, in this work, a thorough literature review of spatial
solar forecasting methods that utilize both satellite imagery and deep
learning is conducted, particularly for large-area applications such as
grid integration. The main contributions of this work are summarized
as follows:

• This work provides a comprehensive summary of spatial solar

forecasting methods that leverage satellite data and deep learn-
ing, a topic not previously reviewed in detail.
• Practical recommendations on end-to-end data-driven forecasting
Fig. 1. Temporal and spatial coverage of commonly employed techniques for obtaining
and essential model development tools are provided, such as
exogenous data.
Source: Modified based on [2,25]. remote-sensing satellite data sources or deep learning implemen-
tation frameworks. This information is beneficial for researchers
to conduct research and model development along this line more
efficiently.
and movement of clouds are highly relevant to surface irradiance
• Existing challenges of spatial solar forecasting are identified,
and thus solar power [27], the inclusion of cloud information de-
providing valuable insights for the development and optimization
rived from satellite imagery as exogenous inputs has been used to
of future methods for hundreds of DSGs across a large spatial area.
enhance forecasting performance [3,28–30]. Images from the latest-
• To overcome the identified challenges, several potential appro-
generation geostationary satellites are currently being sampled every
aches are provided that could potentially improve the perfor-
5 to 15 min, which enables not only intra-day but also intra-hour solar
mance and the utility/practical value of DSG power forecasts.
forecasting, which is particularly useful for power system operations,
These approaches mainly refer to attention mechanisms, sequence
such as load following and real-time electricity trading [30,31]. On
analyses, large-scale models, and spatial probabilistic forecasting
the other hand, the spatial coverage of satellite-based solar forecasts
techniques.
is beneficial for the utility-scale market operators to balance grids
for cities or even states [32,33]. Traditionally, remote-sensing-based In what follows, basic considerations for spatial solar forecasts
solar forecasting is performed under a physical setting, in which the are summarized in Section 2. Spatial solar forecasting methods are
cloud advection and diffusion processes are captured and described [25, reviewed in Sections Section 3. In Section 4, major challenges are iden-
34]. However, statistically determining the cloud dynamics and its tified and an outlook of potential solutions is provided. The conclusions
impact on ground irradiance challenge these physics-based models, are summarized in Section 5. It should be again highlighted that the
often resulting in large forecast errors [35]. Therefore, data-driven term ‘‘solar forecasting’’, whenever used, denotes both solar irradiance
methods, including deep learning, have been introduced to enhance the forecasting and solar power forecasting, not just in this work, but in
remote-sensing-based solar forecasting models [12]. this field in general [43].
Deep learning methods (e.g., deep neural networks, or DNNs) have
brought breakthroughs in many fields such as computer vision and 2. Basic considerations of spatial solar forecasting
speech recognition [36]. Evaluated with various datasets, deep learning
methods have shown superior performance in terms of accuracy, gen- In this section, the data-driven spatial solar forecasting procedure
eralization ability, and robustness, as compared to conventional rule- is firstly presented (Section 2.1). Then the most advanced geosta-
based models or machine learning models [37–40]. The convolutional tionary satellite data around the globe are thoroughly summarized
neural networks (CNNs) and recurrent neural networks (RNNs) are (Section 2.2), followed by common practices to process the satellite
particularly useful in computer vision and time series analysis, respec- data before feeding into the forecasting models (Section 2.3). The
tively [41,42]. With no exception, solar forecasting, as a fast-advancing forecasting performance assessment methods are briefly summarized in
field, has already seen deep learning applications, which have shown Section 2.4.
superior performance over physical or conventional machine learning
methods on many occasions [12,43,44]. 2.1. Overview of data-driven spatial forecasting procedure
However, due to the end-to-end (also known as ‘‘black box’’) na-
ture of data-driven techniques, deep-learning-based solar forecasting The general data-driven spatial solar forecasting procedure is il-
methods are mostly developed using locally collected data as training lustrated in Fig. 2. The procedure can be divided into three blocks,
targets. Therefore, most of these models are developed for a limited including (1) data sources, (2) data processing methods, and (3) fore-
number of point locations and can hardly be transferred or generalized casting model development, assessment, and optimization. The data
to new locations beyond the training set [45]. This is also confined by employed predominantly originate from local sensors, satellite obser-
the fact that deep learning models, unlike statistical regression models vations, and numerical weather prediction (NWP) models. Then, the
(e.g., linear regression), do not have ease-to-express function forms. data are processed by methods such as filtering, normalization, or
So models fitted in one scenario are not conducive to adoption in feature extraction, during which the physical considerations such as
another scenario, unless the code and data are passed on in their exact clear-sky modeling are integrated, to isolate the known (i.e., calculable)
form, which has rarely been the case under the current ‘‘publish or deterministic trends and seasonality in data as far as possible. The time
perish’’ research regime. Consequently, only a few studies have utilized stamps of the processed data from different sources are then matched
both satellite imagery and deep learning to perform spatial/regional to serve as inputs and training targets for forecasting models. Different
forecasting for solar irradiance/power [46,47]. For example, by the algorithms can be employed to perform either single-location forecasts

3
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Fig. 2. Data-driven spatial forecasting procedure.

or spatial forecasts. The forecast accuracy is quantified by different enables the monitoring of different atmospheric constituents, especially
metrics and the preliminary model is further optimized to achieve the the optical and physical properties of clouds (as shown in Fig. 4). The
best performance. Due to the scarcity of local-sensing data over a large comparison of visible, near-infrared, and infrared spectral bands of
area and the recent advancements of geostationary satellite imagers, different satellites are illustrated in Fig. 4 and Table A.2.
remote-sensing data sources have become more applicable for spatial The GOES series has been operated by the National Oceanic and
forecasting applications. Solar irradiance predictions from NWP can Atmospheric Administration, and the National Aeronautics and Space
be used either as model inputs or as validation benchmarks for the Administration of the United States since 1975.
developed forecasting models.
• Imager: Advanced Baseline Imager (ABI).
The spatio-temporal resolution of satellite-based forecasts is intrinsi-
• Level 1 data: the upwelling radiance of each spectral band.
cally tied to the resolution of the satellite data employed—the highest-
• Level 2 data: aerosol detection, aerosol optical depth, clear sky
resolution product has a native resolution of 500 m [48] with most
masks, cloud layers/heights, cloud and moisture imagery, cloud
being 2–4 km. The irradiance prediction for a satellite pixel is applica-
optical depth, cloud particle size distribution, cloud top height,
ble to all DSGs within the coverage area of that pixel. However, such
cloud top phase, cloud top pressure, cloud top temperature, de-
a resolution may be inadequate for forecasts that necessitate a higher
rived motion winds, derived stability indices, downward short-
spatio-temporal resolution, such as intra-hour forecasting [12], par-
wave radiation at surface, fire/hot spot characterization, land sur-
ticularly under partly cloudy conditions where ground irradiance can
face albedo, land surface bidirectional reflectance factor, land sur-
substantially decrease within a minute [49,50]. To accommodate fore-
face temperature (skin), legacy vertical moisture profile, legacy
casting systems requiring higher spatio-temporal resolutions, ground- vertical temperature profile, radiances, rainfall rate, reflected
based sensors or cameras are advantageous for capturing minute-scale shortwave radiation from the top of atmosphere, sea and lake
fluctuations in cloud cover and ground-based irradiance. Subsequent ice age, concentration and motion, sea surface temperature, snow
spatial analysis methods, including inverse distance weighting (IDW), cover, total precipitable water, volcanic ash detection and height
kriging, or nearest neighbor methods, are implemented to derive values [53].
for locations without ground-based sensors. The forecasting perfor- • Data retrieval: open source for non-commercial use and can be
mance of these ground-based systems is notably influenced by the downloaded directly from Amazon Web Services.1
density of the sensor network [11]. However, the installation and
maintenance of a dense sensor network are both costly and resource- European Organization for the Exploitation of Meteorological Satel-
intensive. Consequently, selecting the appropriate data sources and lites (EUMETSAT) is an intergovernmental organization that is mostly
forecasting methods based on application requirements proves to be of funded by European Union countries. The Meteosat satellite series
paramount importance. operated by EUMETSAT provides useful data products for meteorology
and climate applications. Meteosat-1 is the first satellite of the series
2.2. Geostationary satellite data and was launched in November 1977. There are seven satellites of the
first generation of Meteosat and four satellites of the second gener-
Major geostationary satellite data sources that are commonly em- ation of Meteosat, including the currently operating Meteosat-9 and
ployed for solar forecasting are summarized here, including Geostation- Meteosat-11.
ary Operational Environmental Satellites (GOES), Meteosat, Himawari, • Imager: Spinning Enhanced Visible and InfraRed Imager (SEVIRI).
Fengyun, and Chollian. It should be noted that each of these mentioned • Level 1.5 data: the image radiometry with linearization and equal-
names refers to a series of satellites rather than just a single one. The ization of each spectral band.
satellites are operated and managed by different authorities, with dif-
ferent onboard imagers, data resolutions, and coverages, as presented
in Table 1 and Fig. 3. The advanced satellite imagers are capable of 1
https://noaa-goes16.s3.amazonaws.com/index.html, https://noaa-goes17
capturing upwelling radiance data from different spectral bands, which .s3.amazonaws.com/index.html.

4
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table 1
Metadata of different operational geostationary satellites.
Satellite name Position Launch time Number of bands Spatial resolution Observation area and temporal resolution
GOES-16 75.2◦ W Nov, 2016 16 b1–b2: 0.5–1 km Full disk (15 min)
GOES-17a 137.3◦ W Mar, 2018 b3–b6: 1–2 km USA area (5 min)
b7–b16: 2 km Selected areas (0.5 min)
Meteosat-9 45.5◦ E Dec, 2005 12 b1–b3: 1 km Full disk (15 min)
Meteosat-11 0◦ Jul, 2015 b4–b12: 3 km European area (5 min)
Himawari-8 140.7◦ E Oct, 2014 16 b1–b3: 0.5–2 km Full disk (10 min)
b4–b6: 1–2 km Japan area (2.5 min)
b7–b16: 2 km Target area (0.5 min)
Lanmark areas (0.5 min)
Fengyun-4A 104.7◦ E Dec, 2016 14 b1–b2: 0.5–1 km Full disk (15 min)
Fengyun-4B 123.5◦ E Jun, 2021 b3–b6: 2–4 km China area (4–5 min)
b7–b14: 4 km
GK2A 128.2◦ E Dec, 2018 16 b1–b4: 0.5–1 km Full disk (10 min)
b5–b6: 2 km South Korea area (2.5 min)
b7–b16: 2 km
a
GOES-17 is replaced by GOES-18 in 2023, and its position is at 137.0◦ W.

Fig. 3. The spatial observation area of the major operational geostationary satellites. The depicted global annual irradiance map is retrieved from the Solargis website.

Fig. 4. (a) Illustration of different spectral bands of different satellites. (b) Downwelling ground-level solar shortwave (𝜆 < 4 μm) and atmospheric longwave (𝜆 > 4 μm) flux density
for clear and cloudy conditions. The spectral longwave and shortwave flux densities are calculated using the radiative transfer models presented in [51,52], respectively.

5
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

• Level 2 data: aerosol properties over the sea, cloud analysis, cloud Meteorological Satellite (COMS), Geostationary-Korean Multi-purpose
Analysis image, clear-sky reflectance map, divergence, high- Satellite-2A (Geo-KOMPSAT-2A, or GK2A) and Geo-KOMPSAT-2B
resolution precipitation index, normalized difference vegetation (GK2B), all operated by the Korea Aerospace Research Institute.
index/normalized difference vegetation index decadal, tropo-
spheric humidity [54]. • Imager: Advanced Meteorological Imager (AMI).
• Data retrieval: open source and can be downloaded from the • Level 1A data: units of calibrated radiance without navigation.
EUMETSAT website2 for non-commercial use. • Level 1B data: calibrated, navigated, and in-pixel-shape spectral
radiation.
Himawari is a series of geostationary weather satellites launched • Primary meteorological products and meteorological products:
and operated by the Japan Meteorological Agency. Himawari-8 was land surface emissivity, surface albedo, snow depth, sea surface
launched in October 2014 and has been operational since July 2015, current, cloud type, cloud amount, cloud optical depth, cloud ef-
aiming to provide data related to weather forecasting and climate fective radius, cloud liquid water path, cloud layer/height, prob-
change prediction. ability of rainfall, potential accumulated rainfall, angstrom expo-
• Imager: Advanced Himawari Imager (AHI). nent product, visibility, reflected shortwave radiation, downward
• Level 1 data: albedo in solar spectrum, brightness temperature shortwave radiation at surface, absorbed shortwave radiation
in the atmospheric longwave spectrum, satellite zenith angle, at surface, downward longwave radiation at surface, upward
satellite azimuth angle, solar zenith angle, solar azimuth angle, longwave radiation at surface, outgoing longwave radiation, ic-
and observation hour. ing, overshooting top, SO2 detection, total precipitable water,
• Level 2 and level 3 data: aerosol property, sea surface tem- tropopause folding and turbulence detection [57].
perature, nighttime sea surface temperature, shortwave radia- • Data retrieval: open source and can be downloaded from National
tion/photosynthetically available radiation, chlorophyll-a, cloud Meteorological Satellite Center5 (NMSC) for non-commercial use.
property (day-time only) and wildfire [55].
• Data retrieval: open source and can be downloaded from the 2.3. Processing of satellite data
Japan Aerospace Exploration Agency website3 for non-commercial
use. Since the observation areas of a satellite are usually much larger
• Notes: Himawari-8 has both full-disk and regional (i.e., over than the region of interest (ROI) to spatial solar forecasting, such
greater Japan) scans, with the latter being more frequent and as the region covering a cluster of DSG systems connected to the
more refined in resolution than the former. The scanning for the same feeder node, the initial step in the processing of satellite data
target area and landmark areas is determined based on changes involves the selection of an appropriate rectangular region that fully
in weather conditions. encompasses the ROI. To enhance the performance and robustness of
solar forecasting applications, it is essential to pre-process the satellite
The Fengyun series satellites are meteorological satellites devel-
data prior to developing the forecasting model [31].
oped by the Shanghai Institute of Aerospace Technology, China. The
A fundamental and time-tested approach for normalization is the
Fengyun series is mainly used for services such as weather analysis and
minmax method [58], which is a similar but weaker version of the
forecasting, or environmental and disaster monitoring. There are sev-
well-known Heliosat-2 method in this field. It takes the form:
eral pre-determined positions for Fengyun satellites, such as 104.7◦ E, ( )
86.5◦ E, or 123.5◦ E. For Fengyun-4A, the fixed position has been at 𝐿𝑏𝑖𝑗,𝑡 − min 𝐿𝑏
𝐿̃ 𝑏𝑖𝑗,𝑡 = ( ) ( ), (1)
104.7◦ E since 2017. max 𝐿𝑏 − min 𝐿𝑏
• Imager: Advanced Geostationary Radiation Imager (AGRI). where 𝐿𝑏𝑖𝑗,𝑡 is the value of the pixel situated at the 𝑖th-row and 𝑗th-
• Level 1 data: geo-located and calibrated spectral radiance. column in a satellite image of spectral band 𝑏 at time 𝑡. This value
• Level 2 and level 3 data: cloud mask, cloud type, total cloud may represent radiance, albedo, or brightness temperature, contingent
( )
amount, rainfall rate/quantitative precipitation estimate, atmo- on the data product sourced from the specific satellite. min 𝐿𝑏 and
( 𝑏)
spheric motion vector, outgoing longwave radiation, black body max 𝐿 denote the minimum and maximum pixel values across all
brightness temperature, surface solar irradiance, legacy vertical images of band 𝑏 in the entire dataset, respectively.
moisture profile, layer precipitable water, aerosol detection (in- In addition to this standard normalization typically employed in
cluding smoke and dust), sea surface temperature (skin), snow computer vision, seasonality removal should be considered to enhance
cover, land surface (skin) temperature, cloud-top height, cloud- model performance and robustness. Although the utilization of clear-
top pressure, cloud optical depth, cloud liquid water, cloud parti- sky models for eliminating seasonality from solar irradiance time series
cle size distribution, cloud phase, downward longwave radiation has become a standard practice in solar resource assessment and fore-
at surface, upward longwave radiation at surface, reflected short- casting [59], a significant portion of data-driven forecasting models
wave radiation at top of atmosphere, aerosol optical depth, con- in the extant literature neglects to incorporate seasonality removal
vective initiation, fire/hot spot characterization, fog detection, when using satellite data as inputs. This omission leads to a failure to
land surface emissivity, land surface temperature, land surface acknowledge the influence of inherent diurnal and seasonal variations
albedo and tropopause folding turbulence prediction [56]. on the upwelling radiance within the solar spectrum. Ensuring the
• Data retrieval: open source and can be downloaded from the independence of external cloud-related inputs, such as the cloud index
China Meteorological Administration (CMA) website4 with the (CI), from the time of year and season is crucial for enhancing the
requested permit for non-commercial use. robustness of solar forecasting [60]. For instance, high-albedo pixels in
Chollian is the geostationary satellite series launched and operated satellite images, which could be snow or ice, may be incorrectly iden-
by South Korea, which is used for observing oceanography and mete- tified as clouds. Therefore, these upwelling radiance values recorded
orology. The series contains 3 satellites: Communication, Ocean and in satellite data are also subject to inherent diurnal and seasonal
variations, potentially introducing noise into the input data and causing
significant prediction errors.
2
https://www.eumetsat.int/eumetsat-data-store.
3
https://www.eorc.jaxa.jp/ptree/.
4 5
https://satellite.nsmc.org.cn/PortalSite/Data/Satellite.aspx. https://datasvc.nmsc.kma.go.kr/datasvc/html/main/main.do.

6
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

To remove the seasonality in satellite data, Matsunobu et al. [61] can be used to quantify the forecast quality. Some commonly used met-
proposed a method that weighs each satellite spectral band and re- rics are presented in Table A.3, among which a highly recommended
moves diurnal effects (only for bands within the solar spectrum). This scoring function for solar forecasting is the root mean square error
normalization method is mathematically expressed as: (RMSE) [59].
( ) However, forecast accuracies such as RMSE attained from different
𝐿𝑏𝑖𝑗,𝑡 − min 𝐿𝑏
𝐿̃ 𝑏𝑖𝑗,𝑡 = 1 − [ ( ) ( )] (2) forecasting contexts cannot be compared directly, because accuracy
cos 𝜃𝑧,𝑡 max 𝐿𝑏 − min 𝐿𝑏 is highly influenced by factors such as diverse geographic, seasonal,
where 𝜃𝑧,𝑡 denotes the solar zenith angle at time 𝑡. climatic, or particularly meteorological conditions [12,70]. Therefore,
Furthermore, Si et al. [19] proposed an advanced pre-processing reference models often serve as the benchmark to compare the per-
method designed to remove diurnal effects while preserving valuable formance of models proposed in different studies, which use differ-
information. The initial step in this method involves determining the ent datasets for model development and evaluation. The irradiance
coefficient (𝛼) related to the relative positions between the satellite forecasted by NWP is sometimes used as a benchmark for forecasts
and target locations. This is optimally achieved by minimizing the generated by data-driven and/or hybrid methods. The performance of
NWP models has been often found to be outperformed by satellite-based
Euclidean distance between the satellite albedo image and ground
or data-driven methods when the forecast horizons are less than four
albedo images, under the assumption that the ground albedo within the
hours [71,72]. In addition, advanced deep learning methods usually
target area remains constant over a short time period [19]. The value
employ conventional machine learning models as the reference, and
of a single pixel is then normalized by the solar zenith angle as:
hybrid approaches often employ their individual building blocks as the
𝐿𝑏𝑖𝑗,𝑡 reference [3,12]. To offer a relative independent baseline, a once-
𝐿′𝑏
𝑖𝑗,𝑡 = . (3) popular standard of reference is persistence, which assumes that the
(cos 𝜃𝑧,𝑡 )𝛼
current situation persists into the future:
Subsequently, 𝐿′𝑏 ̃𝑏
𝑖𝑗,𝑡 is further normalized to 𝐿𝑖𝑗,𝑡 employing the 𝑚𝑖𝑛𝑚𝑎𝑥
method as presented in Eq. (1). The method presented by Si et al. 𝐼̂𝑡+ℎ = 𝐼𝑡 , (5)
[19] is explicitly designed for visible spectral bands, and its applica-
where 𝐼 could be either solar irradiance or power. 𝐼𝑡 is 𝐼 at current
bility to other bands cannot be assured without further research and
time 𝑡, 𝐼̂ is the forecast of the persistence model, and ℎ is the forecast
experimentation.
horizon. To consider the inherent diurnal cycles in solar irradiance
Additionally, the semi-empirical Heliosat method can be utilized for or power time series, the smart persistence (SP) [73,74] has been
deriving a cloud index map from satellite images as a data processing proposed:
method [48,62–64]. For instance, this can be represented as:
( ) 𝜅̂ 𝑡+ℎ = 𝜅𝑡 , (6)
𝐿𝑏𝑖𝑗,𝑡 − min 𝐿𝑏𝑖𝑗,𝑡−𝑁∶𝑡
CI𝑖𝑗,𝑡 = ( ) ( ) (4) where 𝜅 is the clear-sky index [3], which is defined as solar irradiance
max 𝐿𝑏𝑡 − min 𝐿𝑏𝑖𝑗,𝑡−𝑁∶𝑡 or power divided by its clear-sky counterpart:

where CI𝑖𝑗,𝑡 is the cloud index of the pixel( situated)at the 𝑖-row and 𝜅𝑡 = 𝐼𝑡 ∕𝑐𝑡 , (7)
𝑗-column in a satellite image at time 𝑡. min 𝐿𝑏𝑖𝑗,𝑡−𝑁∶𝑡 is the minimum where 𝑐𝑡 represents the clear-sky expectation of the forecast variable in
pixel value over the preceding 𝑁 days (𝑁 = 10 as per [62]) at the question.
same time, which approximates the corresponding clear sky values, and
( ) Note that the persistence model has shown excellent performance
max 𝐿𝑏𝑡 is the maximum pixel value in the current image. during clear weather conditions. However, the error of the persistence
An alternative approach for removing the diurnal and seasonal vari- model increases with the forecast horizon rapidly. Therefore, the op-
ations inherent in satellite data involves the implementation of cloud timal convex climatology–persistence combination has been advocated
detection. This technique involves the segmentation or identification of by Yang [75,76]:
cloud cover in satellite imagery, which is subsequently utilized as input
for forecasting models. Cloud detection can be achieved through tradi- 𝜅̂ 𝑡+ℎ = 𝛾ℎ 𝜅𝑡 + (1 − 𝛾ℎ )𝜇, (8)
tional pixel color ratio methodologies [65], or through the application where 𝛾ℎ is the autocorrelation at lag ℎ with a value between −1 and
of more sophisticated end-to-end deep learning techniques [61,66–68]. 1, which statistically approaches zero with the increase of forecast
These methods generate binary or three-state cloud maps by denoting horizon. More details about the practical calculation of 𝛾ℎ can be found
pixels as clear skies, thin clouds, or thick clouds. A more comprehensive in [75,76], which is important because the solar time series contains
overview of cloud detection methodologies can be found in [69]. nighttime gaps. 𝜇 is the climatology reference term [76], which is
Notwithstanding, the incorporation of a cloud-detection process com- single-valued internal climatology (i.e., the mean value of 𝜅𝑡 in the
plicates the overarching forecasting system, potentially resulting in verification set). The error of the convex model can be mathematically
extended inference times. Moreover, the accuracy of the cloud detec- shown to be always less than both the persistence and the climatology
tion process itself may be susceptible to the influence of the diurnal and references, thus serving as a better standard of reference [59].
seasonal variations of satellite data. Based on the aforementioned reference models, forecast skill (𝑠)
Despite these various efforts, a majority of the solar forecasting is proposed to quantify the performance of a newly developed model
literature appears to have largely overlooked the importance of pre- when compared with the persistence model [73]:
processing satellite data, particularly in terms of seasonality removal,
RMSE𝑓 − RMSE𝑟
prior to the model estimation phase. 𝑠= . (9)
RMSE𝑝 − RMSE𝑟
2.4. Performance assessment of spatial solar forecasts where the RMSE𝑓 , RMSE𝑝 , RMSE𝑟 are performances of the forecasts
from the model of interest, perfect forecasts, and reference forecasts,
At the current stage of development of the field, the performance respectively. Since perfect forecasts are usually unattainable—except
assessment of spatial forecasts adapts and resembles that of forecasts when dynamical ensemble NWP forecasts are available [77]—one has
for point locations. The assessment is usually performed by comparing no choice but to assume the perfect forecasts have an RMSE of zero. In
the model predictions 𝐼̂ against measured values 𝐼, which are also that, the above skill score equation is reduced to:
referred to as the ground truth or labels. Solar forecasts are continuous RMSE𝑓
in values, and there are many scoring functions (i.e., error metrics) that 𝑠=1− . (10)
RMSE𝑟

7
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Fig. 5. Summary of the reviewed references. (a) Forecast variables, where solar radiation includes global horizontal irradiance (GHI), direct normal irradiance (DNI) and diffuse
horizontal irradiance (DHI). (b) Forecast horizon: For papers with two or more forecast horizons, those are counted separately. (c) Geostationary satellites: All geostationary
satellites used in the literature are listed, and their temporal and spatial resolutions are annotated with points (for a single value) and dash lines (for a range). Points with only
temporal or spatial information are placed on the axes. (d) Input types: Measurements include historical irradiance data and PV output, and satellite products mean the data derived
from satellite measurements. (e) Methods: CNN, long short-term memory (LSTM), graph neural network (GNN) and their combinations are counted as DNN, whereas multilayer
perceptron (MLP) is placed under artificial neural networks (ANN).

In general, a positive 𝑠 indicates the model has more accurate forecasts CNNs are effective in extracting spatial features from images. Ad-
than the reference model, and vice versa. When compared with other ditionally, CNNs are able to analyze spatially correlated data, such as
statistical metrics, 𝑠 is more independent of the forecasting objects multi-dimensional image/video data in computer vision tasks. Com-
(irradiance or power) and climatic or meteorologic factors. Therefore, bining both facts, CNNs align well with the utilization of sky-camera
𝑠 can be used to compare the performance of different models from and satellite images, which act as exogenous inputs facilitating spatial
different works that are developed and validated using different data. solar forecasting. Since processes of solar irradiance or cloud fields
are often spatially correlated, significant enhancements in forecasting
3. Spatial solar forecasting with deep learning techniques performance of both centralized or spatially distributed systems have
been observed [12]. Commonly used CNN architectures include, but are
A literature search was conducted on the Web of Science using the not limited to, AlexNet [81], ResNet [41], DenseNet [82], and Inception
keywords deep learning and spatial solar forecast. A total of 45 publica- net [83]. More details about CNN can be found in [41,84,85], and
tions were found, among which only 20 considered satellite data. These other comprehensive reviews of deep learning methods can be found
references are selected, thoroughly reviewed, and summarized in Fig. 5 in [36,86].
and Table 3. While deep-learning-based methods for solar forecasting have demon-
strated promising results in the literature, particularly due to their
3.1. Overview of deep learning techniques ability to bypass error-prone intermediate physical processes and di-
rectly produce the final prediction, these methods do come with limi-
Deep learning methods, which inherit the structures of ANN [36], tations. First, such models require large amounts of high-quality data
are strong analytic tools for data-driven applications. Deep learning and robust computational resources, which can significantly increase
methods have been widely adopted in a range of applications, as one of overall system costs. As a result, their use is often confined to locations
the most promising state-of-the-art computer science technologies. One with ample local measurements for model development and validation.
of the most established and popularly employed ANNs in both research Second, deep learning models can be prone to overfitting, leading to a
and application is the MLP [12]. For time series analysis, RNNs are decrease in performance when addressing unseen data. For instance,
often found advantageous. For the analysis of spatial information, CNNs predictions pertaining to sudden weather changes that are absent from
are often perceived as more suitable. the training dataset are likely to exhibit substantial errors. Third,
As an effective tool, RNN is capable of processing multiple in- models refined for one location may not perform as well in a different
puts and thus can analyze temporally dynamic behaviors [78]. More location due to variations in local conditions and system configurations.
advanced variants of RNN, such as LSTM [79] or gated recurrent The dynamic nature of city-scale DSGs, including new installations,
unit (GRU) [80], are developed for long time series analysis. More decommissioning, and maintenance events, presents a significant chal-
specifically, compared to basic RNN units, LSTM/GRU units have gates lenge to the adaptive capabilities of data learning models. Last, the
to control information reserving and discarding, which is useful to solve ‘‘black box’’ nature of these models undermines their transparency
the problem of vanishing gradient when the input time sequence is and trustworthiness, a factor that is crucial when understanding the
long. LSTM has hitherto been popular in the domain of solar forecast- decision-making process is important. Consequently, grid operators
ing, but GRU, which has fewer parameters than LSTM, is more friendly may have concerns regarding the explainability of model predictions
towards computational efficiency. and the diagnosability of potential model failures.

8
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table 2
Summary of deep learning frameworks.
Frameworks Initial release Platform Written language Open source Official link
Theano 2007 Cross-platform Python Yes www.deeplearning.net/software/theano
Deeplearning4j 2014 Linux, MacOS, C++, Java Yes www.deeplearning4j.org
Windows, Android
(Cross-platform)
TensorFlow 2015 Linux, MacOS, Python, C++ Yes www.tensorflow.org
Windows,
Android, JavaScript, etc.
Keras 2015 Linux, MacOS, Python Yes keras.io
Windows,
Cross-platform.
MXNet 2015 Linux, MacOS, C++, Python, Yes mxnet.apache.org
Windows R, Java, Julia,
JavaScript, Scala, Go, Perl
Chainer 2015 Linux, MacOS Python Yes chainer.org
PyTorch 2016 Linux, MacOS, Python, C++ Yes pytorch.org
Windows, etc.
CNTK 2016 Linux, Windows C++ Yes www.microsoft.com/en-us/cognitive-toolkit
Caffe 2017 Linux, MacOS, C++ Yes caffe.berkeleyvision.org
Windows
Flux 2017 Linux, MacOS, Julia Yes https://fluxml.ai
Windows
(Cross-platform)
ONNX 2017 Windows, Linux Python, C++ Yes onnx.ai
Matlab Deep – Linux, MacOS, C, C++, Java, No mathworks.com/products/deeplearning
Learning Toolbox Windows MATLAB

3.2. Tools to implement deep learning methods 3.3. Solar forecasting for multiple point locations

Building a deep learning model from scratch is neither efficient For spatial solar forecasting methods using deep learning, early
nor economical. Therefore, most researchers place their attention on works were mostly interested in forecasting separately at point lo-
the specific applications, rather than implementation. On the contrary, cations within an area. The general procedure first trains an MLP
utilizing well-tested deep learning frameworks to develop the model or an extended version of that (i.e., DNN) to provide deterministic
of interest has been found sufficient in most cases. Although other forecasts for point locations [87,88], and then derives the spatial
options are available (see Table 2), the TensorFlow (Keras) and PyTorch forecast maps based on spatial statistics, interpolation, or other space-
frameworks are the most representative and therefore recommended, in filling strategies [11,88–90] (as shown in Fig. 2). Some proposed
consideration of their open-source nature, as well as rich community models are optimized to provide forecasts for multiple locations si-
and documentation support. Indeed, most of the state-of-the-art deep multaneously, which can significantly lower the efforts required for
learning methods and architectures are openly available in these frame- model training and testing [87]. These works mostly employ historical
works. Detailed mathematical theories behind several representative irradiance/power time series and satellite images as default model
deep learning algorithms are presented in Appendix B. In what follows, inputs. To further enhance the model performance, some works em-
a brief account of Tensorflow and PyTorch is given. ploy NWP forecasts and meteorological data, such as temperature and
TensorFlow supports a variety of platforms, such as Linux, Win- relative humidity, as additional exogenous inputs [87,90]. At present,
dows, MacOS, iOS, or Android, and it can work with a range of most methods available in the literature are developed at an hourly
programming languages, such as Python, C++, or GO. TensorFlow is resolution.
particularly popular in commercial and industrial applications due to The aforementioned deep learning models are either in the form
its computational efficiency. TensorFlow is based on a symbolic math of MLP or as hybrids with other methods that integrate conventional
library with excellent documentation on its official website containing machine learning and/or data analysis, such as support vector machine
all modules. Therefore, researchers can customize deep learning models (SVM) [91], principal component analysis, or wavelet transform anal-
with TensorFlow, but its computational graph is not easy to use for ysis [92]. These methods are developed and validated independently
beginners. For beginners, Keras, which is used on top of TensorFlow, is for locations in different regions with different local climates, such as
thought to be the better option. Keras is an open-source high-level deep the Brazilian Northeastern region [89], Italy [90], Netherlands [87],
learning API, written in Python. Keras allows fast experimentation, so Greece [93], Reunion island [92], and South Korea [88,94]. Con-
researchers can quickly prototype and validate their ideas and concepts. sequently, the reported accuracies of these models are not directly
However, Keras is less configurable to build large or new deep learning comparable. When validated using local data, most works suggest that
models. deep-learning-based forecasting methods have superior performance
PyTorch is developed based on Torch, which is mainly used for over conventional methods, such as random forest (RF), support vector
sequence models, reinforcement learning models, computer vision mod- regression (SVR), or unprocessed physical models [88,94].
els, and relational models. It is written in Python, and its API is
similar to other deep learning frameworks like TensorFlow. Similar to 3.4. Advanced approaches to analyze spatial–temporal information
TensorFlow, PyTorch is compatible with all mainstream platforms and
is strongly supported by great documentation and developer commu- To investigate the spatial–temporal correlations of solar irradi-
nity. Whereas both PyTorch and TensorFlow are popular in research ance/power time series measured at different locations, more complex
and development communities, the former has advantages in terms of deep learning architectures are employed for spatial solar forecasting.
flexibility and ease of use. In this section, RNNs for analysis of time series and CNNs for analysis

9
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

of spatial information, particularly for satellite image feature extraction the deep generative framework, which usually has encoder–decoder
and analysis, are to be discussed. architectures [106] where the backbones of the encoder or decoder are
Since the data involved in forecasting, whether as input or output, usually U-net or other CNN variants. These models are able to coopera-
is often time series, it is commonplace to employ an RNN to describe tively process both spatial meteorological data and satellite image data
the input–output relationship. RNNs are capable of predicting the in order to directly generate spatial solar forecasting. Compared with
sequences of solar irradiance or power based on input sequences of other machine learning models, such as ARMA, ARIMA, GRU, LSTM-
historical solar irradiance, satellite and local-sensing images, power fully connected (LSTM-FC), or CNN-bidirectional LSTM (CNN-biLSTM),
measurement, or other meteorological variables. For example, Srivas- these models are reported to have better performance for various
tava and Lessmann [95] used an LSTM model to forecast day-ahead forecast resolution and are capable of serving up to 50 real-world
solar irradiance using satellite data. The proposed model was applied PV stations [13,106]. Other innovative approaches for satellite-based
and validated at 21 locations (16 in Europe and 5 in the United spatial forecasts include, but are not limited to graphical learning
States). The forecasting results suggested that the LSTM model out- framework [72] and extreme learning machine (ELM) [107], which
performed gradient boosting regression (GBR) and feed-forward neural can investigate areas as large as a country (e.g. Australia). To address
networks (FFNN) in terms of day-ahead GHI forecasting. LSTM and the challenge that spatial distributed PV systems have different capac-
other RNN models can also be integrated with conventional machine ities and configurations, a machine learning framework [108], which
learning models to perform solar forecasting. For instance, Kim et al. integrates different data sources to construct a database for regional
[96] applied a stacking ensemble technique to develop a hybrid sea- models, has been proposed to capture the point-to-point relation be-
sonal autoregressive integrated moving average with exogenous factors tween satellite data and in-situ measurements. This method is suggested
(SARIMAX)-LSTM model that used satellite images and local PV power to be site-adaptive to any region where both satellite and ground data
data as inputs. This model analyzed and predicted spatial and temporal are available.
characteristics for PV power in various regions of South Korea. The In addition to statistical forecast accuracy, some works investi-
SARIMAX-LSTM model had superior performance to the benchmark gate the enhancements of other utilization values of forecasts. For
DNN, LSTM, and SARIMAX models. example, most satellite-image-based works are developed for intra-day
Unlike LSTM, CNN is usually combined with a regression (e.g., MLP) forecasting with hourly predicting resolution. To achieve even higher
to conduct solar forecasting [19,97]. The purpose of CNN is to analyze resolutions (e.g. intra-hour forecasting), Cheng et al. [72] proposed a
the satellite images and extract features (e.g., cloud cover information) spatial–temporal graphical learning framework (i.e., GNN) for intra-
hour PV power prediction using historical image sequences as inputs.
from those. Subsequently, these extracted features are combined with
This framework used bi-directional extrapolation to simulate cloud
locally measured irradiance or meteorological data to form the inputs
motion and generated a directed graph to represent the shapes and
to the regression to generate solar forecasts. The CNN-derived feature
motion directions of the regions of interest. It is suggested that GNN
map consists of spatial information, which could be used to estimate
was more flexible for varying sizes of inputs, decreased the number of
a spatially continuous forecast map to cover multiple locations [97].
model inputs and parameters, reduced the data storage requirements,
Most CNN+MLP structures are developed for hourly spatial irradi-
and was more practical for deployment of intra-hour and intra-day PV
ance forecasting. The CNN integrated models have been reported to
power forecast with horizons from 30 min to 3 h. For even higher
achieve improved performance against stand-alone CNN, MLP, and
resolution, Yao et al. [106] proposed a hybrid model that integrated
unprocessed commercial NWP solutions, such as the European Centre
both U-net and encoder–decoder architecture to achieve intra-hour
for Medium-Range Weather Forecasts (ECMWF) [97–99].
forecasting of PV power. This model cooperatively processed both time
The above works employ either CNNs or RNNs independently,
series of historical data and satellite image data to learn both the spatial
whereas methods were also proposed to combine both CNN and RNN
and temporal features. Compared with other machine learning models,
structures, such as CNN+LSTM/GRU, to simultaneously investigate
such as ARMA, ARIMA, GRU, LSTM-FC, or CNN-biLSTM, the hybrid
spatial and temporal information in an end-to-end fashion. Although
model is reported to have the best performance within a 1-h horizon
named differently, the basic structure of these hybrid models resembles at 15-min resolution. On the contrary, some works develop models
those CNN+MLP structures discussed above. The CNN parts of these for much longer time intervals. For example, Deo et al. [107] devel-
models are used to analyze satellite images for feature extraction, while oped an ELM that was regionally adaptive to achieve monthly-scale
the MLP part is replaced by the RNN unit, which is capable of analyzing irradiance forecasting. Using eight Moderate Resolution Imaging Spec-
time series and generating sequential forecasts during the inference troradiometer (MODIS) satellite data variables and geo-topographical
stage. Therefore, satellite images, PV power, irradiance, and other site characteristics as inputs, this model was validated and deployed at
meteorological data can be prepared in the form of time sequences 41 local sites distributed over Australia.
to estimate the model parameters in an end-to-end learning approach. Another approach to enhance the utilization values for spatial solar
Among these hybrid models, most works employ LSTM while only a forecasting is to estimate the uncertainty associated with the prediction
few works [100] employ GRU. map. There are always errors associated with forecasts in practice [109,
Hybrid models have been developed to provide spatial–temporal 110]. Therefore, probability density function forecasts convey more
forecasts for both PV power [14,100] and surface irradiance [21,101– useful information for the decision-making of grid operators [111,112].
103] with forecast horizons ranging from one hour to one day. Some For example, Yang [113] proposed a probabilistic ensemble model
hybrid structures can serve forecasting for up to 400 DSGs with care- output statistics (EMOS) method to provide regional scale solar ir-
fully prepared datasets [100]. Developed and validated in different radiance forecasts for intra-hour horizons. The EMOS model allowed
locations, such as the United States [14], China [100], Brazil [103] and variance scaling and smoothing to optimize the prediction methods.
South Korea [101], hybrid model architectures have shown significant However, to the best knowledge of the authors, very few research
advantages in improving the overall forecasting performance over a has investigated probabilistic spatial solar forecasting. More discussion
set of reference models, such as stand-alone autoregressive integrated about probabilistic solar forecasting for point locations can be found
moving average (ARIMA), MLP, RNN, GRU and LSTM [104,105]. in [12,50].
In general, a hybrid architecture or learning framework, such as hy-
3.5. Other state-of-the-art approaches brid CNN+RNN models or attention-mechanism-based methods, is able
to analyze time series data and generate sequential predictions. These
In the realm of solar forecasting, researchers are progressively hybrid models have demonstrated advantages in improving spatio-
adopting state-of-the-art deep learning methodology, such as inno- temporal forecasting performance and have been reported to outper-
vative learning strategies and application frameworks. For example, form simple machine learning models, such as ARMA, ARIMA, GRU,

10
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

LSTM-FC, or CNN. It is important to note that the performance of configuration information as local inputs (e.g. panel orientations and
these models may vary depending on the specific dataset, location, and shading factor).
climate conditions. Further research and validation are necessary to In recent years, the deep learning research community has been
determine the best-performing model for a given circumstance. rapidly evolving and innovative methods have emerged. Therefore, the
In summary, enhanced with deep learning techniques, spatial fore- employment of more pioneering deep learning methods to analyze the
casting has been proposed in the literature to issue forecast maps spatially distributed solar resource or power is suggested. For example,
for cloud, solar irradiance, and distributed PV power, using satellite the attention mechanism [119] based methods could address the spatial
data with or without local measurements [93]. Related references are solar forecasting problems. The deep-learning-based attention model
summarized and compared in Fig. 5 and Table 3. However, spatial mimics the biological cognitive attention mechanism that focuses on
forecasting for DSG is relatively challenging since individual DSG sites specific aspects of a complex input [120]. As a result, the attention
may have different configurations, such as capacity, efficiency, or track- mechanism divides a complicated problem into smaller and simpler
ing methods. Therefore, available methods usually develop forecast tasks [121]. The attention mechanism has been successfully employed
frameworks that could be quickly adapted to new DSG sites within the in computer vision tasks, such as classification, object detection, se-
investigated regions when locally measured data (e.g. power) becomes mantic segmentation, video understanding, image generation, and 3D
available. Spatial forecasting of cloud and/or solar irradiance is some- vision [122–125]. There are two major types of attention mechanisms
times employed as exogenous inputs to enhance the performance of for computer vision tasks: soft attention and hard attention [126].
these frameworks [114]. Soft attention usually employs a group of filters to create a blurring
effect that the background is faded or blurred while the ROI is in the
4. Outlooks for future research focus. For hard attention, it detects the ROI and discards the entire
background. More information about computer-vision-orientated atten-
The majority of solar forecasting research focused on predicting tion mechanism can be found in [127]. The representative attention
irradiance/power output for a single/point location. However, it is methods are promising for spatial solar forecasting include but are not
challenging to deploy these point-location-oriented forecasting meth- limited to SENet [128], Image GPT [129], Vision Transformer [130],
ods for a large spatial area because: (1) Most available forecasting and Swin Transformer [131].
engines require extensive local-sensing data to develop and validate
Another potential approach is to employ advanced deep learning
the models, particularly for deep learning methods. However, it is very
methods that are capable of video or image sequence forecasts. Com-
costly to deploy a large number of distributed local-sensing systems to
pared to other meteorologic factors such as temperature, pressure,
collect the required data. (2) The relationship between solar irradiance
and aerosol concentrations, volatile cloud fields are the major fac-
and solar power from individual PV systems is highly dependent on
tors that cause intra-hour and intra-day solar ramps [65]. Therefore,
system configurations such as panel orientations, local shading, lo-
the temporal–spatial features of cloud distribution and movement are
cal meteorological conditions, and sun-tracking mechanism [116,117].
useful exogenous inputs for intra-day solar forecasting. Satellite im-
Therefore, for solar power forecasting, the models developed for one
age sequence contains essential information to extract useful cloud
site may not be suitable to be directly applied to other sites. (3) Once
properties and cloud movement features. For example, Lu et al. [132]
a data-driven model is trained and deployed in real-time operation,
proposed a hybrid method that integrated both Cascade Causal LSTM
it may not adapt to unexpected noise or changes in the system or
and Super-Resolution Network to predict the shape and speed of cloud
environment, such as efficiency degeneration due to dirt or aging,
motion based on sky images. For remote-sensing-based methods, Xu
partial system malfunction, and shading from growing trees or new
et al. [133] predicted the next few images of the satellite image
buildings. Currently, most available solar forecasting methods are de-
sequence based on a hybrid of generative adversarial networks (GAN)
veloped based on the assumption that the system will be operating
and LSTM. This work suggested that the GAN-LSTM model is capable of
stably without considering the conditions of partial malfunctions or
combining the generating ability of the GAN with the forecasting ability
unexpected service disruptions. Therefore, more advanced and practical
forecasting methods that are capable of addressing the above issues of the LSTM. Similarly, Rüttgers et al. [134] predicted a typhoon track
should be developed. Here, a few potential approaches for spatial solar using a GAN that analyzes the sequence of satellite images. However,
forecasting are presented in the following with the aim of overcoming how to enhance the aforementioned approaches for spatial solar fore-
the challenges. casts requires further investigation. Therefore, it is suggested to develop
One potential approach is to develop practical and economical end-to-end spatial solar forecasting approaches using the analysis of
solutions that provide full-spectrum forecasts with minimum require- both spatial and temporal features extracted from the sequence of
ments of local-sensing equipment. Although local-sensing systems have satellite images as inputs.
significant advantages in terms of data fidelity and temporal resolu- In addition, probability forecasting is another potential approach
tions, especially for intra-hour and intra-day forecast horizons, instal- to further assist solar integration [12]. Most available solar forecast-
lation, maintenance, data synchronizing, and preprocessing costs of ing methods, either for the point location or the large areas, predict
local-sensing systems will scale up dramatically with the number of deterministic values. However, there are always inherent and irre-
sites of interest. In addition, deployment of local-sensing systems usu- ducible errors associated with deterministic predictions regardless of
ally requires complicated procedures to obtain hardware installation the data processing, training methods, and model mechanisms [109,
permits due to considerations including but not limited to security, 110]. Therefore, the quantified solar forecast uncertainty is useful
safety, and weatherproof. Therefore, with the advancement of new information for grid operators to integrate solar power while mini-
generations of geostationary satellites, utilizing remote-sensing satellite mizing the adversary impacts from forecast errors. The methods to
data is potentially an alternative approach to deliver intra-hour and estimate solar forecast uncertainty include but not limited to delta
intra-day solar forecasts for numerous DSGs over large areas. Yagli techniques [135], Bayesian methods [112], bootstrap method [136],
et al. [118] investigated the performance difference between ground- bootstrap-ANN method [137], quantiles methods [138,139], lower up-
based and satellite-based forecasting models. The results suggested that per bound estimation method [110], kNN ensemble model [50], naïve
satellite-based forecasts have comparable performance to ground-based Bayes classifier and Bayesian models [140,141]. More details of prob-
forecasts. Therefore, with the advancement of deep learning and spatial abilistic forecasting approaches can be found in [142–144]. Here,
interpolation techniques, forecast engines using only exogenous inputs employing the aforementioned methods for spatial solar forecasting and
such as satellite data may be proposed. The advanced methods are generating spatial distributed forecast uncertainty map is suggested for
expected to be quickly adapted to any location with simple system future work.

11
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table 3
Summary of the publications of spatial solar irradiance or power forecasting based on remote sensing and deep learning.
Authors Year Forecast Temporal and Input local-sensing data Input remote-sensing data Methods
variables and spatial resolution of
horizons satellite data
Lima et al. [89] 2016 24-h-ahead GHI Not used Not used 2 years of hourly GHI Not used Weather Research and
measurement data of 110 Forecasting (WRF) and ANN
weather stations in
northeastern Brazil
Pierro et al. [90] 2017 0 to 48-h-ahead Not Not 2 years of PV output data 2 years of satellite-derived ANN, referenced by smart
PVoutput provided provided from 11 sites in the south irradiance data from persistent model
of Italy and irradiance METEOSAT-10
data forecast by WRF-AWR
Lago et al. [87] 2018 0 to 6-h-ahead Not Not 4 years of daily ECMWF 4 years of satellite-derived DNN, referenced by
GHI provided provided forecasts, Ineichen–Perez irradiance data from Royal persistence model, linear
clear-sky irradiance Netherlands Meteorological model, extreme gradient
Institute boosting (XGBoost) and
ECMWF forecasts
Srivastava and 2018 0 to 24-h-ahead Not Not Not used 10 years of satellite-based LSTM, referenced by FFNN,
Lessmann [95] GHI provided provided GHI data from 21 stations GBR and persistence model
in Europe and US
Zhang et al. [91] 2019 0 to Not used Not used 1 year of PV output data Not used Bayesian Network, referenced
30-min-ahead PV and 6 types of by ARIMA, 𝑘 nearest neighbor
output meteorological data from (kNN), LSTM, persistent model
two open data portals and spatio-temporal baseline
methods
Yeom et al. [94] 2019 GHI estimation 1 h 1 to 4 km 6 years of 35 ground 6 years of satellite images Physical model, ANN, RF, SVR
pyranometers from Korea from COMS and DNN, RF and DNN show
Meteorological better results
Administration (KMA)
Nikitidou et al. 2019 0 to 6 h-ahead 15 min 5.5 km Not used 3 years of satellite-derived ANN
[93] cloud clearness CCI from METEOSAT
index (CCI) Second Generation
Jeong and Kim 2019 0 to 6-h-ahead Not used Not used 1 year of PV output from Not used CNN, referenced by AR,
[14] PV output National Renewable Energy FFNN, and LSTM
Laboratory in the USA
Liu et al. [100] 2019 3-h-ahead GHI Not used Not used 2 years of hourly GHI and Not used Variational Bayesian
meteorological data from convolutional GRU, referenced
National Solar Radiation by simple DNN, RNN and
Database (NSRDB) LSTM
Jiang et al. [98] 2019 GHI estimation Not Not 2 years of hourly GHI of 2 years of satellite images DNN, referenced by DNN with
provided provided 98 ground irradiance from Himawari-6 less inputs and ANN
stations from CMA
Khodayar et al. 2019 0 to 2-h-ahead Not used Not used 18 years of GHI data in 75 Not used Convolutional graph
[13] GHI sites near Lake Michigan autoencoder, referenced by
from NSRDB space–time Copula,
spatio-temporal lasso quantile
regression, Compressive
Spatio-temporal Forecasting
and spatio-temporal SVR
Doorga et al. 2019 0 to 5-day-ahead Not 3.3 km 3 years of solar irradiance 3 years of satellite data ARMA, nonlinear
[115] GHI provided in the island of Mauritius from METEOSAT First autoregressive neural network
Generation and double exponential
smoothing (DES), DES has
better results
Deo et al. [107] 2019 GHI Not Not 4 year of GHI from 4 years of satellite-derived ELM (originated from ANN),
provided provided Scientific Information for data from MODIS Terra referenced by RF, M5 Tree,
Land Owners in Australia sensor multivariate adaptive
regression spline
Yang [113] 2020 GHI estimation Not used Not used 6 satellite-derived Not used EMOS technique
databases and 2global
reanalysis databases
Li et al. [92] 2020 GHI estimation Not used Not used 18 years of global solar Not used ANN
irradiance data from CM
SAF
Kim et al. [88] 2020 1 to 2-h-ahead 15 min 0.5 to 2 1 year data of PV power 1 year of satellite images ANN, DNN and SVM, DNN
PV output km output from SK Telecom from Himawari-8 and has better results
COMS

(continued on next page)

12
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table 3 (continued).
Authors Year Forecast Temporal and Input local-sensing data Input remote-sensing data Methods
variables and spatial resolution of
horizons satellite data
Si et al. [19] 2020 0 to 4-h-ahead 1 h 5 km 2 years of hourly GHI data 2 years of satellite images CNN+ANN, referenced by very
GHI and meteorological data in from Fengyun-2G deep convolutional networks,
Shandong from CMA LSTM+ANN
Pérez et al. [97] 2020 0 to 6-h-ahead 15 min 3 km 2.5 years of GHI data in 2.5 years of satellite DNN, referenced by smart
GHI the southern part of France images from METEOSAT persistent model
from a pyranometer Second Generation
Yeom et al. 2020 GHI estimation 15 min to 1 to 4 km 4.75 years of GHI 4.75 years of satellite ConvLSTM, referenced by RF
[101] and 1-h-ahead 3 h measurement from 33 images from COMS and ANN
forecast ground station from KMA
in Korea
Hong et al. 2020 0 to 24-h-ahead 1 h Not 1 year of GHI data from 1 year of satellite-derived ConvLSTM, referenced by
[102] GHI provided the SolarGIS database GHI data CNN-LSTM, ARIMA and LSTM
Kim et al. [96] 2021 0 to 1-h-ahead 15 min 2 km 4 years of PV output data 4 years of satellite images SARIMAX, SVR, LSTM, DNN,
PV output in South Korea from Open from COMS RF and SARIMAX-LSTM,
Data Porta, meteorological SARIMAX-LSTM shows better
data results
Narvaez et al. 2021 GHI estimation 30 min 111 km 10 years of 30 min GHI, 10 years of satellite data RF for estimation, LTSM+GRU
[108] and 24 to DNI, DHI and from NSRDB for forecast
168-h-ahead GHI meteorological data in
forecasts Columbia, South America
Yao et al. [106] 2021 0 to 1-h-ahead 10 min Not 9 months of PV output 9 months of satellite CNN+LSTM+AM, referenced
PV output provided data from 50 PV stations, images from Himawari-8 by ARMA, ARIMA, GRU,
NWP products, LSTM-FC, CNN-biLSTM
meteorological data
Cheng et al. [72] 2021 0.5 to 3-h-ahead 10 min 5 km 2 years of PV output from 2 years of satellite images GNN, referenced by smart
PV output DKA, clear-sky irradiance, from Himawari-8 persistence, deep belief
meteorological data network, ConvLSTM, 3D-CNN,
and 2D-MC CNN
Chu et al. [11] 2022 GHI, DNI and Not used Not used 2 months of images from Not used Linear model for DHI, MLP
DHI estimation sky cameras in California for GHI and DNI, and kriging
for spatial GHI
Jang et al. [99] 2022 GHI estimation 2 to 0.5 to 1 year of GHI 1 year of satellite images CNNs
10 min 2 km measurement from 81 from GK2A
ASOS stations
Qin et al. [21] 2022 0–6-h-ahead GHI 1 h 5 km 1 year of GHI 1 year of satellite images CNN+LSTM, referenced by
measurement from 10 from Himawari-6 CLSTM, ConvLSTM and smart
ground stations in China persistence
Rocha and 2022 GHI and DNI 15 min 0.5 to GHI, DNI DHI Satellite images from XGBoost and CNN-LSTM
Santos [103] estimation 2 km measurement data from GOES-16
SONDA station in Brazil

Since 2021, large-scale models have emerged in leading journals In summary, there are still gaps between state-of-the-art spatial
for weather forecasting based on satellite imagery [145,146]. These solar forecasting methods and grid demands, particularly in the era of
pioneering techniques provide promising opportunities for developing massive deployment of distributed solar power techniques in different
innovative forecasting strategies that can effectively support a wide regions. In addition to solely focus on statistical accuracy, the applica-
range of real-world DSG operations. For instance Pangu-Weather, a tion scenarios and associated grid operational policy should be carefully
deep learning-driven weather forecasting system, utilizes reanalysis analyzed to identify the optimal forecasting approach. In the future,
weather data to predict future weather conditions. It employs a 3D research of solar forecasting shall be more application orientated, cost-
Earth-specific transformer architecture and hierarchical temporal ag- effective, and practical to be deployed in order to satisfy the immediate
gregation, enhancing accuracy while reducing computational time. needs of power integration.
Tested on ERA5 data, this system has demonstrated notable improve-
ments in accuracy compared to other models [147]. Similarly, the
5. Conclusion
FourCastNet model, a data-centric, high-resolution weather forecast-
ing solution, is engineered to predict variables such as surface winds
This review focuses on the current state-of-the-art spatial solar
and precipitation. It offers superior resolution compared to traditional
models and delivers performance comparable with physics-based mod- forecasting methods, particularly those that employ deep learning and
els [148]. Other large models for precipitation forecasting are also remote-sensing techniques, with a special emphasis on grid-scale DSGs
discussed in the literature [145,149]. Despite these advancements, over large areas. Fundamental considerations for spatial solar forecast-
there remain substantial research gaps in the application of these mod- ing, including data-driven forecasting procedures, satellite data sources
els to solar forecasting, which has higher variability when compared and processing, tools for implementing deep learning methods, and
with other weather variables. These issues include low data quality performance assessments, are provided in a comprehensive summary
and availability, high computational resource requirements and high to facilitate further research. Subsequently, the available spatial solar
spatio-temporal resolution requirements of solar forecasting. As such, forecasting methods, their inputs, and associated deep learning models
the adaptability of these weather-oriented large-scale models to predict are reviewed, compared, and discussed.
DSG generations with diverse parameters still need further research At present, forecasting solutions for utility-scale DSGs are still in the
efforts. early stages of research and implementation. There has been limited

13
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

investigation into this field due to various challenges, such as high mandatory assessment metrics [3]. Zhang et al. [155] evaluated statis-
scale-up costs, limited flexibility for adaptation to real-life operations, tical metrics based on Western Wind and Solar Integration Study Phase
and the complexity involved in considering the diverse configurations 2 data [156] and suggested that most of these metrics are effective in
of installed DSGs. However, advances in remote-sensing techniques and assessing uniform forecast improvements. Therefore, the assessment of
deep learning methods have led to the identification of several promis- solar forecasts is application-oriented that based on local regulations
ing approaches that could enhance spatial solar forecasting solutions. and application/operation requirements.
There is a need for more research efforts to develop next-generation
adaptive spatial solar forecasting methodologies, which should be capa- Appendix B. Deep learning algorithms
ble of meeting both the technological and commercial demands posed
by the substantial integration of DSG systems. Deep learning methods [36] are strong analytic tools for data-driven
learning, which have been widely applied for a range of applications as
CRediT authorship contribution statement one of the most promising state-of-the-art technologies. Here, the math-
ematical theories behind several representative deep learning methods
Yinghao Chu: Conceptualization, Methodology, Formal analysis, are presented. More comprehensive review of deep learning methods
Validation, Writing – original draft. Yiling Wang: Investigation, Writ- can be found in [36,86].
ing – original draft. Dazhi Yang: Validation, Investigation, Writing –
review & editing. Shanlin Chen: Visualization, Writing – review & B.1. Artificial neural networks (ANNs)
editing. Mengying Li: Methodology, Software, Validation, Investiga-
tion, Resources, Writing – review & editing, Visualization, Supervision, ANNs are inspired by the biological neural systems [159]. Compu-
Project administration, Funding acquisition. tations of each ANN neuron, which is responsible for signal processing
tasks, are mathematically defined as
Declaration of competing interest (𝑁 )
∑
𝑦 = 𝜎𝐴 𝑤𝑖 𝑥𝑖 + 𝑤0 , (B.1)
The authors declare that they have no known competing finan- 𝑖=1
cial interests or personal relationships that could have appeared to where 𝑥 is the neuron input and 𝑦 is the neuron output, 𝑤𝑖 and
influence the work reported in this paper. 𝑤0 are the weights and bias, 𝜎𝐴 is the activation function, such as
sigmoid, hyperbolic tangent, rectified linear unit (ReLU) and Leaky
Data availability ReLU functions [160,161]. The widely used activation functions and
corresponding derivatives are summarized in Table B.4, while more
Data will be made available on request. detailed discussion of deep learning activation functions can be found
in [160,162]. One of the most popularly employed ANNs in both
Declaration of Generative AI and AI-assisted technologies in the research and application is the MLP [12]. The MLP neurons are placed
writing process in layers, and neurons in one layer are fully forward connected with
neurons in the following layer. The outputs of neurons in one layer are
During the revision process of this work, the authors used GPT-4 in used as the inputs for neurons in the following layer.
order to improve the language. After using this tool/service, the authors
reviewed and edited the content as needed and take full responsibility B.2. ANN parameter estimation
for the content of the publication.
The weights and bias of ANNs are mostly estimated using back prop-
Acknowledgments agation learning [163,164], which is an supervised learning method.
The general procedure of back propagation training is:
Yinghao Chu is substantially supported by a grant from City Univer-
sity of Hong Kong (Project No. 9610625). Mengying Li is substantially 1. Set the number of layers 𝑁𝐿 and the number of neurons 𝑁𝑁 for
supported by a grant from the Research Grants Council of Hong Kong each layer to define the architecture of the ANN.
(Project No. 25213022). 2. Set the training parameters, such as tolerance parameter 𝑒𝑇 >0,
the learning rate 𝛼𝐿 , and the maximum epochs of data learning.
Appendix A. Supplemental information 3. Initialize the ANN weight vectors 𝑤𝑙𝑘 𝑘 for all layers 𝑙 =
𝑙 𝑙−1
1, 2, … , 𝑁𝐿 and all neurons 𝑘𝑙 in layer 𝑙 and 𝑘𝑙−1 in layer 𝑙 −
Commonly employed NWP models for solar forecasting include the 1. Initialization can be done randomly or by using pre-trained
ECMWF [16], the global forecast system (GFS) [150], the Hirlam– weights from another dataset.
Aladin research towards mesoscale operational NWP in Euromed (HAR- 4. For each (input vector 𝑥,) compute the output of each neuron
∑ 𝑙
MONIE) model, the rapid refresh (RAP) model, the North American 𝑦𝑙𝑘 = 𝜎𝐴 𝑙−1
𝑛 𝑤𝑘𝑙 𝑘𝑙−1 𝑦𝑘𝑙−1 for all layers 𝑙 and all neurons 𝑘𝑙 in
𝑙
mesoscale (NAM) model, the high resolution rapid refresh (HRRR) layer 𝑙.
model, the WRF model and its solar parameterization [151,152], and 5. Compute the loss 𝜀 of the current iterative training epoch. The
the global and regional assimilation and prediction system - global loss function could be the mean square error or cross-entropy
forecast system (GRAPES_GFS) [153]. The comparison of different NWP loss between the network outputs and the training targets.
models are presented in Table A.1, and the start, central and end 6. Compute the derivative of the cost function with respect to the
wavelengths (in μm) of each band of different satellite imagers are output of each neuron in the output layer, then compute the
𝑁 𝑁 𝑁
compared in Table A.2. error 𝛿𝑘 𝐿 for each output neuron 𝑘𝑁𝐿 as: 𝛿𝑘 𝐿 = 𝜕𝜀∕𝜕𝑦𝑘 𝐿 𝜎𝐴′
The commonly used accuracy evaluation metrics are summarized ( )𝑁𝐿 𝑁𝐿 𝑁𝐿
𝑁
in Table A.3. The European and International Energy Agency recom- 𝑧𝑘 𝐿 , where 𝑧 represents the input of the specific neuron.
𝑁𝐿
mends the use of mean biased error (MBE), RMSE, and Kolmogorov– 7. For each layer 𝑙 = 𝑁𝐿 − 1, 𝑁𝐿 − 2, … , 1, compute the error 𝛿𝑘𝑙 for
Smirnov integral (KSI) to assess the solar forecast performance [154]. (∑ ) ( ) 𝑙
𝑙+1 𝑙+1
Industrial/utility regulations frequently require relative errors such as each neuron 𝑘𝑙 as: 𝛿𝑘𝑙 = 𝑘𝑙+1 𝑤𝑘 𝑘 𝛿𝑘 𝜎 ′ 𝑧𝑙𝑘 .
𝑙 𝑙 𝑙−1 𝑙+1 𝑙
relative mean absolute error (rMAE) and relative RMSE (rRMSE) as 8. Update the neuron weights by: 𝑤𝑙𝑘 𝑘 = 𝑤𝑙𝑘 𝑘 − 𝛼𝐿 𝛿𝑘𝑙 𝑦𝑙−1
𝑘
.
𝑙 𝑙−1 𝑙 𝑙−1 𝑙 𝑙−1

14
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table A.1
Comparison of different NWP models.
Name Spatial Atmo-sphere Update Forecast Forecast Covered region Main output variables
resolution [km] layers frequency [h] horizon [h] resolution [h]
ECMWF(HRES) 9 137 6 90 1 Global GHI, DNI, extraterrestrial irradiance, surface
pressure, total column water vapor, total
column ozone, forecast albedo, cloud cover,
and air temperature
GFS 13 127 1 120/240 1/3 Global Air temperature, wind, precipitation, soil
/384 /12 moisture, and atmospheric ozone
concentration
HARMONIE 2.5 65 6 48 1 Europe GHI, DNI, air temperatures, relative
humidity, cloud cover, pressure, and
precipitable water
RAP 13 51 1 18 1 North American Wind, precipitation, air temperature and
cloud cover
NAM 12 60 1/3 84 1/3 North American GHI, DNI, DHI and air temperature
HRRR 3 51 1 18/48 0.25/1 North American Air temperature, geopotential height, dew
point temperature, graupel, cloud mixing
ratio and cloud ice
WRF-Solar 1–36 30–50 1–6 72–96 1–6 Regional GHI, DNI, and DHI
GRAPES_GFS 28 60 6 240 6 Global Double-moment cloud physics, cloud
macrophysics and prognostic cloud

Table A.2
Start, central and end wavelengths (in μm) of each spectral band of different satellite imagers. Here, VIS denotes visible, NIR demotes near infrared and IR demotes infrared.
GOES-16 ABI Meteosat-11 SEVIRI Himawari-8 AHI Fengyun-4 AGRI GK2A AMI
Start Central End Start Central End Start Central End Start Central End Start Central End
0.45 0.47 0.49 0.60 0.75 0.90 0.43 0.47 0.48 0.45 0.47 0.49 0.43 0.46 0.48
0.59 0.64 0.69 0.56 0.63 0.71 0.50 0.51 0.52 0.55 0.65 0.75 0.50 0.51 0.52
VIS
0.74 0.81 0.88 0.63 0.64 0.66 0.63 0.64 0.66
0.85 0.86 0.87
0.85 0.86 0.89 1.50 1.64 1.78 0.85 0.86 0.87 0.75 0.83 0.90 1.37 1.38 1.38
1.37 1.38 1.39 1.60 1.60 1.62 1.36 1.38 1.39 1.60 1.61 1.62
NIR 1.58 1.61 1.64 2.25 2.30 2.27 1.58 1.61 1.64
2.23 2.26 2.28 2.10 2.25 2.35
3.80 3.90 4.00 3.48 3.92 4.36 3.74 3.90 3.96 3.50 3.75 4.00 3.74 3.85 3.96
5.77 6.15 6.60 5.35 6.25 7.15 6.06 6.24 6.43 3.50 3.75 4.00 6.06 6.24 6.43
6.75 7.00 7.15 6.85 7.38 7.85 6.89 6.94 7.01 5.80 6.25 6.70 6.89 6.95 7.01
7.24 7.40 7.44 8.30 8.70 9.10 7.26 7.35 7.43 6.90 7.10 7.30 7.26 7.35 7.43
8.30 8.50 8.70 9.38 9.66 9.94 8.44 8.59 8.76 8.00 8.50 9.00 8.44 8.60 8.76
IR 9.42 9.70 9.80 9.80 10.80 11.80 9.54 9.64 9.72 10.30 10.70 11.30 9.54 9.63 9.72
10.10 10.30 10.60 11.00 12.00 13.00 10.30 10.41 10.60 11.50 12.00 12.50 10.25 10.43 10.61
10.80 11.20 11.60 12.40 13.40 14.40 11.10 11.24 11.30 13.20 13.50 13.80 11.08 11.20 11.32
11.80 12.30 12.80 12.20 12.38 12.50 12.15 12.30 12.45
13.00 13.30 13.60 13.20 13.28 13.40 13.21 13.30 13.39
a
Although the visible spectrum is from 0.4 μm to 0.7 μm, and the near-infrared spectrum is from 0.7 μm to 2.5 μm, the classification for visible bands is obtained from NMSC).
b
The band 7 and band 8 of Fengyun-4 have the same spectral range, but the spatial resolution of band 7 is 2 km and that of band 8 is 4 km. Band 7 is for fire observation
while band 8 is for land surface observation [56].

Table A.3
Statistical metrics for performance assessment of solar forecasts.
Definition Notes
∑𝑛 ( ̂ )
Mean bias error (MBE) MBE = 𝑛−1 𝑖=1
𝐼𝑖 − 𝐼𝑖
∑ 𝑛
Mean absolute error (MAE) MAE = 𝑛−1 𝑖=1 |𝐼̂𝑖 − 𝐼𝑖 |
√
∑𝑛 ( )2
Root mean square error (RMSE) RMSE = 𝑛−1 𝑖=1 𝐼̂𝑖 − 𝐼𝑖
( −1 ∑𝑛 )
Relative mean absolute error (rMAE) rMAE = MAE∕ 𝑛 𝑖=1 𝑖
𝐼 The denominator can also be averaged irradiance, peak nominal irradiance, or clear sky index
( ∑𝑛 )
Relative root mean square error (rRMSE) rRMSE = RMSE∕ 𝑛−1 𝑖=1 𝐼𝑖 The denominator can also be averaged irradiance, peak nominal irradiance, or clear sky index
( )
Coefficient of determination (𝑅2 ) 𝑅2 = 1 − Var 𝑰̂ − 𝑰 ∕Var (𝑰)
√
( )
Correlation coefficient (𝜌) 𝜌 = Cov 𝑰,̂ 𝑰 ∕ Var(𝑰)Var(𝑰)
̂

Kolmogorov–Smirnov integral (KSI) KSI = ∫ 𝐷𝐼 𝑑𝐼 𝐷𝐼 is the discrepancy in cumulative distributions between the predictions and the
measurements
√
∑𝑛 ( )2
Error standard deviation (𝜎) 𝜎= 𝑛−1 𝑖=1 𝜀𝑖 − 𝜇 Where 𝜀𝑖 = 𝐼̂𝑖 − 𝐼𝑖 , and 𝜇 is the mean value of 𝜀𝑖
(∑ )
𝑚
Rényi entropy (𝐻𝑎 ) [157] 𝐻𝑎 = (1 − 𝑎)−1 log2 𝑎
𝑗=1 𝑝𝑗 𝑝𝑗 is the probability density for 𝑗th section of the error distribution, 𝑎 is the order of 𝐻𝑎 , and
higher magnitude of 𝑎 puts higher weight on more probable events [158]

15
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

Table B.4
Summary of activation functions of ANNs.
Expression Derivative
{ {
𝑧, if 𝑧≥0 1, if 𝑧 ≥ 0
ReLU 𝜎𝐴 (𝑧) = 𝛿(𝑧) =
0, if 𝑧<0 0, if 𝑧 < 0
{ {
𝑧, if 𝑧 ≥ 0 1, if 𝑧 ≥ 0
Leaky ReLU 𝜎𝐴 (𝑧) = 𝛿(𝑧) =
−𝜖𝑧, if 𝑧 < 0 𝜖, if 𝑧 < 0
1
Sigmoid 𝜎𝐴 (𝑧) = 𝛿(𝑧) = 𝜎𝐴 (𝑧)(1 − 𝜎𝐴 (𝑧))
1 + 𝑒−𝑧
Hyperbolic tangent 𝜎𝐴 (𝑧) = tanh(𝑧) 𝛿(𝑧) = (1 − |𝜎𝐴 (𝑧)2 |)
Gaussian radial basis 𝜎𝐴 (𝑧) = exp(−||𝑧 − 𝑚||2 ∕𝜎 2 ) 𝛿(𝑧) = −2(𝑧 − 𝑚)𝜎𝐴 (𝑧)∕𝜎 2
{ {
1, if 𝑧 > 0 0, if 𝑧 ≠ 0
Unipolar step 𝜎𝐴 (𝑧) = 𝐻(𝑧) = 𝛿(𝑧) =
0, if 𝑧 < 0 ∞, if 𝑧 = 0
{
0, if 𝑧 ≠ 0
Bipolar step 𝜎𝐴 (𝑧) = sign(𝑧) = 2𝐻(𝑧) − 1 𝛿(𝑧) =
∞, if 𝑧 = 0
⎧0, if 𝑧 < −1
⎪
Unipolar linear 𝜎𝐴 (𝑧) = ⎨(𝑧 + 1)∕2, if |𝑧| < 1 𝛿(𝑧) = [𝐻(𝑧 + 1) − 𝐻(𝑧 − 1)]∕2
⎪
⎩1, if 𝑧 > 1
⎧−1, if 𝑧 < −1
⎪
Bipolar linear 𝜎𝐴 (𝑧) = ⎨𝑧, if |𝑧| < 1 𝛿(𝑧) = 𝐻(𝑧 + 1) − 𝐻(𝑧 − 1)
⎪
⎩1, if 𝑧 > 1

9. Repeat the steps 4 to 8 until convergence is achieved, i.e. the of multiple features from the same location on an input image. This
preset maximum epochs has been reached or 𝜀𝑗 −𝜀𝑗−1 ≤ 𝑒T where process can be mathematically defined as [176]:
𝑗 denotes the epoch number.
𝑦𝑘 = 𝜎𝐴 (𝑤𝑘 ∗ 𝑥), (B.2)
10. Output the trained ANN with the final weights.
where 𝑥 represents the input image, 𝑦𝑘 denotes the image features
The initialization of the weights and bias can employ random
extracted using the convolutional layer, 𝜎𝐴 (⋅) is a nonlinear activation
method, the Xavier method [165], or the HE method [166]. Another
function, 𝑤𝑘 is the trainable parameters of the 𝑘th feature extractor
commonly used parameters initialization methods is fine-tuning with
of the convolutional layer, and ∗ represents the convolutional opera-
pretrained weights [167]. First, an ANN is pretrained on a large dataset
tion. Convolutional layers are proposed based on the assumption that
to optimize its parameters. Second, the pretrained parameters will be
different parts of an image may share similar image features or rep-
used as the initialization parameters for new tasks [168,169]. Then,
resentations, allowing the same feature extractor to analyze different
the ANNs will be trained with the training data from new tasks, which
positions on an image [42].
could have a relatively small learning rate.
Pooling layers are often grouped with convolutional layers to form
The training or fine tuning process of ANN can be enhanced using
modules [176]. These modules are typically stacked to form a deep
training optimizers. Adaptive moment estimation algorithm (ADAM)
structure for image feature extraction [177]. A pooling layer is mathe-
[170] is one of the most commonly used optimizer, which is employed
matically defined as a nonlinear down-sampling method, which calcu-
as the default choice for many tasks. More details of other training
lates the maximum or average of features from the same local area [81,
optimizers can be found in [3,159,171,172].
83]. In practice, a Max Pooling Layer, which forwards the maximum
value of the receptive field to the next layer, is more commonly
B.3. Spatial data analysis based on CNNs used. Pooling layers offer several benefits, including the reduction of
computational efforts, mitigation of overfitting, and enhancement of
CNN stands as one of the most effective deep learning tools for ana- spatial invariance [178].
lyzing spatially correlated data, such as multi-dimensional image/video The modules of convolutional and pooling layers generate a high-
data in computer vision tasks. As such, CNNs are commonly em- dimensional matrix of image features. These image features are further
ployed in solar forecasting methods for sky/satellite image analysis, analyzed by the classifier layers to produce the final outputs. Common
significantly enhancing the forecasting performance of both central- classifiers include SVM, MLP, or global pooling layers [179]. MLPs,
ized and spatially distributed systems [12]. The concept of CNNs was which include fully connected layers, are commonly employed due to
first proposed in the 1980s [173]. Training these networks efficiently their capability for arbitrary nonlinear mappings and their flexibility in
was a significant challenge of that era. However, this problem was real-life application scenarios [3]. The number of neurons in the final
later addressed with the introduction of the gradient backpropaga- output layer usually equals the number of classification labels in the
tion algorithm proposed by Al-Saffar et al. [174]. Since then, CNNs training dataset. Further details on MLPs can be found in Appendix B.2.
have demonstrated remarkable performance on datasets such as the
handwritten digit dataset MNIST [175]. Several CNN architectures are B.4. Time series analysis with RNNs
frequently used, including but not limited to AlexNet [81], ResNet [41],
DenseNet [82], and inception net [83]. For more comprehensive in- RNNs can process multiple inputs, enabling the analysis of tempo-
formation about CNNs, readers are directed to the works of [41,84, rally dynamic behaviors [78]. RNNs have been effectively employed to
85]. model and predict solar irradiance or power output time series. The
Most CNN architectures comprise convolutional, pooling, and clas- basic RNN unit is mathematically defined as:
sifier layers [36]. Convolutional layers examine input images using
receptive fields, which function as feature extractors to derive image 𝑎𝑡 = 𝜎𝐴 (𝑤𝑎𝑎 𝑎𝑡−1 + 𝑤𝑎𝑥 𝑥𝑡 + 𝛽𝑎 ), (B.3)
features or representations from different portions of an image. These
feature extractors, with their distinct weights, allow the extraction 𝑦𝑡 = 𝜎𝐴 (𝑤𝑦𝑎 𝑎𝑡 + 𝛽𝑦 ), (B.4)

16
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

where 𝑥𝑡 represents the inputs, 𝑦𝑡 is the output, and 𝑎𝑡 denotes the [11] Chu Y, Li M, Pedro HTC, Coimbra CFM. A network of sky imagers for spatial
hidden state activation at input sequence step 𝑡. 𝜎𝐴 is the activa- solar irradiance assessment. Renew Energy 2022;187:1009–19.
[12] Chu Y, Li M, Coimbra CFM, Feng D, Wang H. Intra-hour irradiance forecasting
tion function. The 𝑤𝑎𝑎 , 𝑤𝑎𝑥 , 𝑤𝑦𝑎 , 𝛽𝑎 , and 𝛽𝑦 are trainable weights or
techniques for solar power integration: A review. iScience 2021;24:103136.
parameters. [13] Khodayar M, Mohammadi S, Khodayar ME, Wang J, Liu G. Convolutional graph
LSTM [79] and GRU [80] are variants of RNN, developed to address autoencoder: A generative deep neural network for probabilistic spatio-temporal
the issue of vanishing gradient when the input time sequence is long. solar irradiance forecasting. IEEE Trans Sustain Energy 2019;11:571–83.
LSTM units, compared with basic RNN units, have an input gate (𝑖𝑡 ), a [14] Jeong J, Kim H. Multi-site photovoltaic forecasting exploiting space–time
convolutional neural network. Energies 2019;12:4490.
forget gate (𝑓𝑡 ), and an output gate (𝑜𝑡 ). New information can be stored
[15] Gao H, Liu M. Short-term solar irradiance prediction from sky images with
in the cell state using 𝑖𝑡 , existing information can be discarded from the a clear sky model. In: Proceedings of the IEEE/CVF winter conference on
cell state using 𝑓𝑡 , and the final output of the LSTM unit is determined applications of computer vision. 2022, p. 2475–83.
using 𝑜𝑡 . The mathematical definitions of these three gates are: [16] Yang D, Wang W, Hong T. A historical weather forecast dataset from the
European Centre for Medium-Range Weather Forecasts (ECMWF) for energy
𝑖𝑡 = 𝜎𝐴 (𝑤𝑖 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝛽𝑖 ), (B.5) forecasting. Sol Energy 2022;232:263–74.
[17] Mayer MJ, Yang D. Probabilistic photovoltaic power forecasting us-
ing a calibrated ensemble of model chains. Renew Sustain Energy Rev
𝑓𝑡 = 𝜎𝐴 (𝑤𝑓 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝛽𝑓 ), (B.6)
2022;168:112821.
[18] Agoua XG, Girard R, Kariniotakis G. Probabilistic models for spatio-temporal
𝑜𝑡 = 𝜎𝐴 (𝑤𝑜 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝛽𝑜 ), (B.7) photovoltaic power forecasting. IEEE Trans Sustain Energy 2018;10:780–9.
[19] Si Z, Yu Y, Yang M, Li P. Hybrid solar forecasting method using satellite
where 𝑤 and 𝛽 represent the weights and biases of the three gates, visible images and modified convolutional neural networks. IEEE Trans Ind
respectively, and ℎ𝑡−1 is the hidden state output from the previous time Appl 2020;57:5–16.
step. For the next time step, the memory state 𝑚𝑡 , candidate cell state [20] Benamrou B, Ouardouz M, Allaouzi I, Ben Ahmed M. A proposed model to
forecast hourly global solar irradiation based on satellite derived data, deep
𝑚̃ 𝑡 , and the output ℎ𝑡 are derived as:
learning and machine learning approaches. J Ecol Eng 2020;21.
[21] Qin J, Jiang H, Lu N, Yao L, Zhou C. Enhancing solar PV output forecast by
𝑚̃ 𝑡 = tanh(𝑤𝑐 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝛽𝑐 ), (B.8)
integrating ground and satellite observations with deep learning. Renew Sustain
Energy Rev 2022;167:112680.
𝑚𝑡 = 𝑓𝑡 𝑚𝑡−1 + 𝑖𝑡 𝑚̃ 𝑡 , (B.9) [22] Wang W, Yang D, Hong T, Kleissl J. An archived dataset from the ECMWF
ensemble prediction system for probabilistic solar power forecasting. Sol Energy
ℎ𝑡 = 𝑜𝑡 tanh(𝑚𝑡 ), (B.10) 2022;248:64–75.
[23] Yang D, Wang W, Bright JM, Voyant C, Notton G, Zhang G, Lyu C. Verifying
where tanh is the hyperbolic tangent function. operational intra-day solar forecasts from ECMWF and NOAA. Sol Energy
2022;236:743–55.
For computational efficiency, GRU has fewer parameters than LSTM.
[24] Jimenez PA, Hacker JP, Dudhia J, Haupt SE, Ruiz-Arias JA, Gueymard CA,
Instead of using three gates, GRU employs an update gate 𝑧𝑡 , a reset Thompson G, Eidhammer T, Deng A. WRF-Solar: Description and clear-sky
gate 𝑟𝑡 , and a memory content ℎ̃𝑡 : assessment of an augmented NWP model for solar power prediction. Bull Am
Meteorol Soc 2016;97:1249–64.
𝑧𝑡 = 𝜎𝐴 (𝑤𝑧 [ℎ𝑡−1 , 𝑥𝑡 ]), (B.11) [25] Yang Dazhi, Kleissl Jan. Solar Irradiance and Photovoltaic Power Forecasting.
CRC Press; 2024.
𝑟𝑡 = 𝜎𝐴 (𝑤𝑟 [ℎ𝑡−1 , 𝑥𝑡 ]), (B.12) [26] Yang D, Perez R. Can we gauge forecasts using satellite-derived solar
irradiance? J Renew Sustain Energy 2019;11:023704.
[27] Yagli GM, Yang D, Srinivasan D. Automatic hourly solar forecasting using
ℎ̃𝑡 = tanh(𝑤ℎ [𝑟𝑡 ℎ𝑡−1 , 𝑥𝑡 ]), (B.13) machine learning models. Renew Sustain Energy Rev 2019;105:487–98.
[28] Zagouras A, Pedro HTC, Coimbra CFM. Clustering the solar resource for grid
With 𝑧𝑡 , 𝑟𝑡 , and ℎ̃𝑡 , the final GRU output is: management in island mode. Sol Energy 2014;110:507–18.
[29] Taravat A, Del Frate F, Cornaro C, Vergari S. Neural networks and support
ℎ𝑡 = (1 − 𝑧𝑡 )ℎ𝑡−1 + 𝑧𝑡 ℎ̃𝑡 , (B.14) vector machine algorithms for automatic cloud classification of whole-sky
ground-based images. Geosci Remote Sens Lett IEEE 2015;12:666–70.
References [30] Pedro HTC, Larson DP, Coimbra CFM. A comprehensive dataset for the
accelerated development and benchmarking of solar forecasting methods. J
[1] Bouckaert S, Pales AF, McGlade C, Remme U, Wanner B, Varro L, D’Ambrosio D, Renew Sustain Energy 2019;11:036102.
Spencer T. Net zero by 2050: a roadmap for the global energy sector. Technical [31] Larson DP, Coimbra CFM. Direct power output forecasts from remote sensing
report, International Energy Agency; 2021. image processing. J Sol Energy Eng 2018;140.
[2] Yang D, Wang W, Gueymard CA, Hong T, Kleissl J, Huang J, Perez MJ, [32] Kaur A, Nonnenmacher L, Pedro HTC, Coimbra CFM. Benefits of solar
Perez R, Bright JM, Xia X, et al. A review of solar forecasting, its dependence forecasting for energy imbalance markets. Renew Energy 2016;86:819–30.
on atmospheric sciences and implications for grid integration: Towards carbon [33] Catalina A, Alaíz CM, Dorronsoro JR. Combining numerical weather predictions
neutrality. Renew Sustain Energy Rev 2022;161:112348. and satellite data for PV energy nowcasting. IEEE Trans Sustain Energy
[3] Inman RH, Pedro HTC, Coimbra CFM. Solar forecasting methods for renewable 2019;11:1930–7.
energy integration. Prog Energy Combust Sci 2013;39:535–76. [34] Kleissl J. Solar Energy Forecasting and Resource Assessment. Academic Press;
[4] Widén J, Carpman N, Castellucci V, Lingfors D, Olauson J, Remouit F, 2013.
Bergkvist M, Grabbe M, Waters R. Variability assessment and forecasting of [35] Chu Y, Urquhart B, Gohari SMI, Pedro HTC, Kleissl J, Coimbra CFM. Short-
renewables: A review for solar, wind, wave and tidal resources. Renew Sustain term reforecasting of power output from a 48 MWe solar PV plant. Sol Energy
Energy Rev 2015;44:356–75. 2015;112:68–77.
[5] Yang D, Wang W, Xia X. A concise overview on solar resource assessment [36] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436–44.
and forecasting. Adv Atmos Sci 2022;39:1239–51. http://dx.doi.org/10.1007/ [37] Weimer D, Scholz-Reiter B, Shpitalni M. Design of deep convolutional neural
s00376-021-1372-8. network architectures for automated feature extraction in industrial inspection.
[6] Lave M, Kleissl J. Solar variability of four sites across the state of Colorado. CIRP Ann 2016;65:417–20.
Renew Energy 2010;35:2867–73. [38] Ren R, Hung T, Tan KC. A generic deep-learning-based approach for automated
[7] Xin-gang Z, Zhen W. Technology, cost, economic performance of distributed surface inspection. IEEE Trans Cybern 2017;48:929–40.
photovoltaic industry in China. Renew Sustain Energy Rev 2019;110:53–64. [39] Masci J, Meier U, Ciresan D, Schmidhuber J, Fricout G. Steel defect classifica-
[8] Zakeri B, Gissey GC, Dodds PE, Subkhankulova D. Centralized vs. distributed tion with max-pooling convolutional neural networks. In: Proc. int. joint conf.
energy storage–benefits for residential users. Energy 2021;236:121443. neural netw.. IJCNN, 2012, p. 1–6.
[9] Jiang S, Wan C, Chen C, Cao E, Song Y. Distributed photovoltaic generation [40] Park J-K, Kwon B-K, Park J-H, Kang D-J. Machine learning-based imaging
in the electricity market: Status, mode and strategy. CSEE J Power Energy Syst system for surface defect inspection. Int J Precis Eng Manuf-Green Technol
2018;4:263–72. 2016;3:303–10.
[10] Chu Y, Pedro HTC, Kaur A, Kleissl J, Coimbra CFM. Net load forecasts for [41] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In:
solar-integrated operational grid feeders. Sol Energy 2017;158:236–46. Proc. IEEE comput. vis. pattern recognit.. CVPR, 2016, p. 770–8.

17
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

[42] Al-Saffar AAM, Tao H, Talab MA. Review of deep convolution neural network [71] Voyant C, Notton G, Kalogirou S, Nivet M-L, Paoli C, Motte F, Fouilloy A.
in image classification. In: 2017 international conference on radar, antenna, Machine learning methods for solar radiation forecasting: A review. Renew
microwave, electronics, and telecommunications. ICRAMET, IEEE; 2017, p. Energy 2017;105:569–82.
26–31. [72] Cheng L, Zang H, Wei Z, Ding T, Sun G. Solar power prediction based on
[43] Yang D, Kleissl J, Gueymard CA, Pedro HTC, Coimbra CFM. History and trends satellite measurements-a graphical learning method for tracking cloud motion.
in solar irradiance and PV power forecasting: A preliminary assessment and IEEE Trans Power Syst 2021.
review using text mining. Sol Energy 2018;168:60–101. [73] Marquez R, Coimbra CFM. Intra-hour DNI forecasting methodology based on
[44] Anagnostos D, Schmidt T, Cavadias S, Soudris D, Poortmans J, Catthoor F. cloud tracking image analysis. Sol Energy 2013;91:327–36.
A method for detailed, short-term energy yield forecasting of photovoltaic [74] Chu Y, Pedro HTC, Coimbra CFM. Hybrid intra-hour DNI forecasts with sky
installations. Renew Energy 2019;130:122–9. image processing enhanced by stochastic learning. Sol Energy 2013;98:592–603.
[45] Liu C, Li M, Yu Y, Wu Z, Gong H, Cheng F. A review of multi-temporal and [75] Yang D. Making reference solar forecasts with climatology, persistence, and
multi-spatial scales photovoltaic forecasting methods. IEEE Access 2022. their optimal convex combination. Sol Energy 2019;193:981–5.
[46] Zhu XX, Tuia D, Mou L, Xia G-S, Zhang L, Xu F, Fraundorfer F. Deep learning [76] Yang D. Standard of reference in operational day-ahead deterministic solar
in remote sensing: A comprehensive review and list of resources. IEEE Geosci forecasting. J Renew Sustain Energy 2019;11:053702.
Remote Sens Mag 2017;5:8–36. [77] Liu B, Yang D, Mayer MJ, Coimbra CF, Kleissl J, Kay M, Wang W, Bright JM,
[47] Carneiro TC, de Carvalho PCM, Alves dos Santos H, Lima MAFB, d. S. Braga AP. Xia X, Lv X, Srinivasan D, Wu Y, Beyer HG, Yagli GM, Shen Y. Predictability
Review on photovoltaic power and solar resource forecasting: current status and and forecast skill of solar irradiance over the contiguous United States. Renew
trends. J Sol Energy Eng 2022;144. Sustain Energy Rev 2023;182:113359.
[48] Huang C, Shi H, Yang D, Gao L, Zhang P, Fu D, Xia X, Chen Q, Yuan Y, [78] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by
Liu M, Hu B, Lin K, Li X. Retrieval of sub-kilometer resolution solar irradiance back-propagating errors. Nature 1986;323:533–6.
from Fengyun-4A satellite using a region-adapted Heliosat-2 method. Sol Energy [79] Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput
2023;264:112038. 1997;9:1735–80.
[49] Chu Y, Li M, Coimbra CFM. Sun-tracking imaging system for intra-hour DNI [80] Medsker L, Jain LC. Recurrent neural networks: design and applications. CRC
forecasts. Renew Energy 2016;96:792–9. Press; 1999.
[50] Chu Y, Coimbra CFM. Short-term probabilistic forecasts for direct normal [81] Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convo-
irradiance. Renew Energy 2017;101:526–36. lutional neural networks. In: Proc. adv. neural inform. process. syst.(neurIPS).
[51] Li M, Liao Z, Coimbra CFM. Spectral model for clear sky atmospheric longwave 2012, p. 1097–105.
radiation. J Quant Spectrosc Radiat Transfer 2018;209:196–211. [82] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected
[52] Li M, Liao Z, Coimbra CFM. Spectral solar irradiance on inclined surfaces: A convolutional networks. In: Proc. IEEE comput. vis. pattern recognit.. CVPR,
fast Monte Carlo approach. J Renew Sustain Energy 2020;12:053705. 2017, p. 4700–8.
[53] Schmit TJ, Griffith P, Gunshor MM, Daniels JM, Goodman SJ, Lebair WJ. [83] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D,
A closer look at the ABI on the GOES-R series. Bull Am Meteorol Soc Vanhoucke V, Rabinovich A. Going deeper with convolutions. In: Proc. IEEE
2017;98:681–98. comput. vis. pattern recognit.. CVPR, 2015, p. 1–9.
[54] Schmetz J, Pili P, Tjemkes S, Just D, Kerkmann J, Rota S, Ratier A. An [84] Zeiler MD, Fergus R. Visualizing and understanding convolutional networks. In:
introduction to Meteosat second generation (MSG). Bull Am Meteorol Soc Proc. europ. conf. comp. vis.. ECCV, 2014, p. 818–33.
2002;83:977–92. [85] Chen CW. Internet of video things: Next-generation IoT with visual sensors.
[55] Bessho K, Date K, Hayashi M, Ikeda A, Imai T, Inoue H, Kumagai Y, IEEE Internet Things J 2020;7:6676–85. http://dx.doi.org/10.1109/JIOT.2020.
Miyakawa T, Murata H, Ohno T, et al. An introduction to Himawari-8/9— 3005727.
Japan’s new-generation geostationary meteorological satellites. J Meteorol Soc [86] Ahmed R, Sreeram V, Mishra Y, Arif M. A review and evaluation of the state-
Jpn Ser II 2016;94:151–83. of-the-art in PV solar power forecasting: Techniques and optimization. Renew
[56] Yang J, Zhang Z, Wei C, Lu F, Guo Q. Introducing the new generation of Sustain Energy Rev 2020;124:109792.
Chinese geostationary weather satellites, Fengyun-4. Bull Am Meteorol Soc [87] Lago J, De Brabandere K, De Ridder F, De Schutter B. Short-term forecasting
2017;98:1637–58. of solar irradiance without local telemetry: A generalized model using satellite
[57] Kim D, Gu M, Oh T-H, Kim E-K, Yang H-J. Introduction of the advanced meteo- data. Sol Energy 2018;173:566–77.
rological imager of Geo-Kompsat-2a: In-orbit tests and performance validation. [88] Kim M, Song H, Kim Y. Direct short-term forecast of photovoltaic power through
Remote Sens 2021;13:1303. a comparative study between COMS and himawari-8 meteorological satellite
[58] Jo J-M. Effectiveness of normalization pre-processing of big data to the machine images in a deep neural network. Remote Sens 2020;12:2357.
learning performance. J Korea Inst Electron Commun Sci 2019;14:547–52. [89] Lima FJ, Martins FR, Pereira EB, Lorenz E, Heinemann D. Forecast for surface
[59] Yang D, Alessandrini S, Antonanzas J, Antonanzas-Torres F, Badescu V, solar irradiance at the Brazilian northeastern region using NWP model and
Beyer HG, Blaga R, Boland J, Bright JM, Coimbra CFM, et al. Verification of artificial neural networks. Renew Energy 2016;87:807–18.
deterministic solar forecasts. Sol Energy 2020;210:20–37. [90] Pierro M, De Felice M, Maggioni E, Moser D, Perotto A, Spada F, Cornaro C.
[60] Ineichen P, Perez R. Derivation of cloud index from geostationary satellites and Data-driven upscaling methods for regional photovoltaic power estimation and
application to the production of solar irradiance and daylight illuminance data. forecast using satellite and numerical weather prediction data. Sol Energy
Theor Appl Climatol 1999;64:119–30. 2017;158:1026–38.
[61] Matsunobu LM, Pedro HTC, Coimbra CFM. Cloud detection using convolutional [91] Zhang R, Ma H, Hua W, Saha TK, Zhou X. Data-driven photovoltaic generation
neural networks on remote sensing images. Sol Energy 2021;230:1020–32. forecasting based on a Bayesian network with spatial–temporal correlation
[62] Paletta Q, Arbod G, Lasenby J. Omnivision forecasting: Combining satellite and analysis. IEEE Trans Ind Inf 2019;16:1635–44.
sky images for improved deterministic and probabilistic intra-hour solar energy [92] Li P, Bessafi M, Morel B, Chabriat J-P, Delsaut M, Li Q. Daily surface solar
predictions. Appl Energy 2023;336:120818. radiation prediction mapping using artificial neural network: the case study of
[63] Rigollier C, Lefèvre M, Wald L. The method Heliosat-2 for deriving shortwave Reunion Island. J Solar Energy Eng 2020;142.
solar radiation from satellite images. Sol Energy 2004;77:159–69. [93] Nikitidou E, Zagouras A, Salamalikis V, Kazantzidis A. Short-term cloudiness
[64] Perez R, Ineichen P, Moore K, Kmiecik M, Chain C, George R, Vignola F. A new forecasting for solar energy purposes in Greece, based on satellite-derived
operational model for satellite-derived irradiances: Description and validation. information. Meteorol Atmos Phys 2019;131:175–82.
Sol Energy 2002;73:307–17. [94] Yeom J-M, Park S, Chae T, Kim J-Y, Lee CS. Spatial assessment of solar radiation
[65] Chu Y, Pedro HTC, Nonnenmacher L, Inman RH, Liao Z, Coimbra CFM. A smart by machine learning and deep neural network models using data provided
image-based cloud detection system for intra-hour solar irradiance forecasts. J by the COMS MI geostationary satellite: A case study in south Korea. Sensors
Atmos Ocean Technol 2014;31:1995–2007. 2019;19:2082.
[66] Berthomier L, Pradel B, Perez L. Cloud cover nowcasting with deep learning. [95] Srivastava S, Lessmann S. A comparative study of LSTM neural networks
In: 2020 tenth international conference on image processing theory, tools and in forecasting day-ahead global horizontal irradiance with satellite data. Sol
applications. IPTA, IEEE; 2020, p. 1–6. Energy 2018;162:232–47.
[67] Wang Z, Zhao J, Zhang R, Li Z, Lin Q, Wang X. UATNet: U-shape attention- [96] Kim B, Suh D, Otto M-O, Huh J-S. A novel hybrid spatio-temporal forecasting
based transformer net for meteorological satellite cloud recognition. Remote of multisite solar photovoltaic generation. Remote Sens 2021;13:2605.
Sens 2021;14:104. [97] Pérez E, Pérez J, Segarra-Tamarit J, Beltran H. A deep learning model for
[68] Roy R, Ahan M, Soni V, Chittora A. Towards automatic transformer-based cloud intra-day forecasting of solar irradiance using satellite-based estimations in the
classification and segmentation. In: NeurIPS 2021 workshop on tackling climate vicinity of a PV power plant. Sol Energy 2021;218:652–60.
change with machine learning. Vol. 2021, 2021, p. 60. [98] Jiang H, Lu N, Qin J, Tang W, Yao L. A deep learning algorithm to estimate
[69] Li L, Li X, Jiang L, Su X, Chen F. A review on deep learning techniques hourly global solar radiation from geostationary satellite data. Renew Sustain
for cloud detection methodologies and challenges. Signal Image Video Process Energy Rev 2019;114:109327.
2021;15:1527–35. [99] Jang J-C, Sohn E-H, Park K-H. Estimating hourly surface solar irradiance from
[70] Marquez R, Coimbra CFM. Proposed metric for evaluation of solar forecasting GK2A/AMI data using machine learning approach around Korea. Remote Sens
models. ASME J Sol Energy Eng 2013;135:0110161–9. 2022;14:1840.

18
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

[100] Liu Y, Qin H, Zhang Z, Pei S, Wang C, Yu X, Jiang Z, Zhou J. Ensem- [128] Hu J, Shen L, Sun G. Squeeze-and-excitation networks. In: Proceedings of the
ble spatiotemporal forecasting of solar irradiation using variational Bayesian IEEE conference on computer vision and pattern recognition. 2018, p. 7132–41.
convolutional gate recurrent unit network. Appl Energy 2019;253:113596. [129] Chen M, Radford A, Child R, Wu J, Jun H, Luan D, Sutskever I. Generative
[101] Yeom J-M, Deo RC, Adamowski JF, Park S, Lee C-S. Spatial mapping of short- pretraining from pixels. In: International conference on machine learning.
term solar radiation prediction incorporating geostationary satellite images PMLR; 2020, p. 1691–703.
coupled with deep convolutional LSTM networks for South Korea. Environ Res [130] Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T,
Lett 2020;15:094025. Dehghani M, Minderer M, Heigold G, Gelly S, et al. An image is worth
[102] Hong Y-Y, Martinez JJF, Fajardo AC. Day-ahead solar irradiation forecasting 16x16 words: Transformers for image recognition at scale. 2020, arXiv preprint
utilizing gramian angular field and convolutional long short-term memory. IEEE arXiv:2010.11929.
Access 2020;8:18741–53. [131] Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B. Swin transformer:
[103] Rocha PA, Santos VO. Global horizontal and direct normal solar irradiance Hierarchical vision transformer using shifted windows. In: Proceedings of the
modeling by the machine learning methods XGBoost and deep neural networks IEEE/CVF international conference on computer vision. 2021, p. 10012–22.
with CNN-LSTM layers: a case study using the GOES-16 satellite imagery. Int [132] Lu Z, Wang Z, Li X, Zhang J. A method of ground-based cloud motion predict:
J Energy Environ Eng 2022;1–16. CCLSTM+ SR-net. Remote Sens 2021;13:3876.
[104] Oh M, Kim CK, Kim B, Yun C, Kang Y-H, Kim H-G. Spatiotemporal opti- [133] Xu Z, Du J, Wang J, Jiang C, Ren Y. Satellite image prediction relying on GAN
mization for short-term solar forecasting based on satellite imagery. Energies and LSTM neural networks. In: ICC 2019-2019 IEEE international conference
2021;14:2216. on communications. ICC, IEEE; 2019, p. 1–6.
[105] Kumari P, Toshniwal D. Deep learning models for solar irradiance forecasting:
[134] Rüttgers M, Lee S, Jeon S, You D. Prediction of a typhoon track using a
A comprehensive review. J Clean Prod 2021;318:128566.
generative adversarial network and satellite images. Sci Rep 2019;9:1–15.
[106] Yao T, Wang J, Wu H, Zhang P, Li S, Xu K, Liu X, Chi X. Intra-hour photovoltaic
[135] Lu T, Viljanen M. Prediction of indoor temperature and relative humid-
generation forecasting based on multi-source data and deep learning methods.
ity using neural network models: model comparison. Neural Comput Appl
IEEE Trans Sustain Energy 2021;13:607–18.
2009;18:345–57.
[107] Deo RC, Şhin M, Adamowski JF, Mi J. Universally deployable extreme learning
[136] Dybowski R, Roberts S. Confidence intervals and prediction intervals for
machines integrated with remotely sensed MODIS satellite predictors over
feed-forward neural networks. Clin Appl Artif Neural Netw 2001;298–326.
Australia to forecast global solar radiation: A new approach. Renew Sustain
[137] Chu Y, Li M, Pedro HTC, Coimbra CFM. Real-time prediction intervals for
Energy Rev 2019;104:235–61.
intra-hour DNI forecasts. Renew Energy 2015;83:234–44.
[108] Narvaez G, Giraldo LF, Bressan M, Pantoja A. Machine learning for
site-adaptation and solar radiation forecasting. Renew Energy 2021;167:333–42. [138] Bremnes JB. Probabilistic wind power forecasts using local quantile regression.
[109] Carney JG, Cunningham P, Bhagwan U. Confidence and prediction intervals for Wind Energy 2004;7:47–54.
neural network ensembles. In: Neural networks, 1999. iJCNN’99. International [139] Bremnes JB. A comparison of a few statistical models for making quantile wind
joint conference on. Vol. 2, IEEE; 1999, p. 1215–8. power forecasts. Wind Energy 2006;9:3–11.
[110] Khosravi A, Nahavandi S, Creighton D. Prediction intervals for short-term wind [140] Nam S, Hur J. Probabilistic forecasting model of solar power outputs based on
farm power generation forecasts. IEEE Trans Sustain Energy 2013;4:602–10. the naive Bayes classifier and kriging models. Energies 2018;11:2982.
[111] Pinson P, Nielsen HA, Møller JK, Madsen H, Kariniotakis GN. Non-parametric [141] Doubleday K, Jascourt S, Kleiber W, Hodge B-M. Probabilistic solar power
probabilistic forecasts of wind power: required properties and evaluation. Wind forecasting using Bayesian model averaging. IEEE Trans Sustain Energy
Energy 2007;10:497–516. 2020;12:325–37.
[112] Bracale A, Caramia P, Carpinelli G, Di Fazio AR, Ferruzzi G. A Bayesian method [142] Pedro HTC, Coimbra CFM, David M, Lauret P. Assessment of machine learning
for short-term probabilistic forecasting of photovoltaic generation in smart grid techniques for deterministic and probabilistic intra-hour solar forecasts. Renew
operation and control. Energies 2013;6:733–47. Energy 2018;123:191–203.
[113] Yang D. Ensemble model output statistics as a probabilistic site-adaptation tool [143] Lauret P, David M, Pinson P. Verification of solar irradiance probabilistic
for satellite-derived and reanalysis solar irradiance. J Renew Sustain Energy forecasts. Sol Energy 2019;194:254–71.
2020;12:016102. [144] Doubleday K, Hernandez VVS, Hodge B-M. Benchmark probabilistic solar
[114] Wang P, van Westrhenen R, Meirink JF, van der Veen S, Knap W. Surface forecasts: Characteristics and recommendations. Sol Energy 2020;206:52–67.
solar radiation forecasts by advecting cloud physical properties derived from [145] Ravuri S, Lenc K, Willson M, Kangin D, Lam R, Mirowski P, Fitzsimons M,
Meteosat Second Generation observations. Sol Energy 2019;177:47–58. Athanassiadou M, Kashem S, Madge S, et al. Skilful precipitation nowcasting
[115] Doorga JRS, Dhurmea KR, Rughooputh S, Boojhawon R. Forecasting mesoscale using deep generative models of radar. Nature 2021;597:672–7.
distribution of surface solar irradiation using a proposed hybrid approach [146] Burke M, Driscoll A, Lobell DB, Ermon S. Using satellite imagery to understand
combining satellite remote sensing and time series models. Renew Sustain and promote sustainable development. Science 2021;371:eabe8628.
Energy Rev 2019;104:69–85. [147] Bi K, Xie L, Zhang H, Chen X, Gu X, Tian Q. Accurate medium-range global
[116] Khoo YS, Nobre A, Malhotra R, Yang D, Ruther R, Reindl T, Aberle AG. weather forecasting with 3D neural networks. Nature 2023;619:533–8.
Optimal orientation and tilt angle for maximizing in-plane solar irradiation for [148] Kurth T, Subramanian S, Harrington P, Pathak J, Mardani M, Hall D, Miele A,
pv applications in singapore. IEEE J Photovolt 2014;4:647–53. Kashinath K, Anandkumar A. FourCastNet: Accelerating global high-resolution
[117] Yoshida S, Ueno S, Kataoka N, Takakura H, Minemoto T. Estimation of global weather forecasting using adaptive fourier neural operators. In: Proceedings of
tilted irradiance and output energy using meteorological data and performance the platform for advanced scientific computing conference. 2023, p. 1–11.
of photovoltaic modules. Sol Energy 2013;93:90–9. [149] Zhang Y, Long M, Chen K, Xing L, Jin R, Jordan MI, Wang J. Skilful nowcasting
[118] Yagli GM, Yang D, Gandhi O, Srinivasan D. Can we justify producing univariate
of extreme precipitation with NowcastNet. Nature 2023;619:526–32.
machine-learning forecasts with satellite-derived solar irradiance? Appl Energy
[150] NOAA. The global forecast system (GFS). 2023, Available at: https://www.emc.
2020;259:114122.
ncep.noaa.gov/emc/pages/numerical_forecast_systems/gfs/documentation.php,
[119] Niu Z, Zhong G, Yu H. A review on the attention mechanism of deep learning.
Accessed on 2023-11-08.
Neurocomputing 2021;452:48–62.
[151] Incecik S, Sakarya S, Tilev S, Kahraman A, Aksoy B, Caliskan E, Topcu S,
[120] Zhao B, Feng J, Wu X, Yan S. A survey on deep learning-based fine-
Kahya C, Odman MT. Evaluation of WRF parameterizations for global horizontal
grained object classification and semantic segmentation. Int J Autom Comput
irradiation forecasts: A study for Turkey. Atmósfera 2019;32:143–58.
2017;14:119–35.
[152] Gueymard C, Jimenez P. Validation of real-time solar irradiance simulations
[121] Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser Ł,
over Kuwait using WRF-solar. In: 12th international conference on solar energy
Polosukhin I. Attention is all you need. In: Advances in neural information
for buildings and industry, euroSun2018. Rapperswil, Switzerland; 2018, p.
processing systems. 2017, p. 5998–6008.
0130.
[122] Wang F, Jiang M, Qian C, Yang S, Li C, Zhang H, Wang X, Tang X.
Residual attention network for image classification. In: Proceedings of the IEEE [153] Shen X, Wang J, Li Z, Chen D, Gong J. Research and operational development
conference on computer vision and pattern recognition. 2017, p. 3156–64. of numerical weather prediction in China. J Meteorol Res 2020;34:675–98.
[123] Mnih V, Heess N, Graves A, et al. Recurrent models of visual attention. In: [154] IEA. Solar heating and cooling programme, task 36, subtask a – standard
Advances in neural information processing systems. 2014, p. 2204–12. qualification for solar resource products. Technical report, IEA; 2012.
[124] Itti L, Koch C. Computational modelling of visual attention. Nat Rev Neurosci [155] Zhang J, A F, Hodge B, Lu S, Hamann H, Banunarayanan V, Brockway A. A
2001;2:194–203. suite of metrics for assessing the performance of solar power forecasting. Sol
[125] Chen L-Q, Xie X, Fan X, Ma W-Y, Zhang H-J, Zhou H-Q. A visual attention Energy 2015;111:157–75.
model for adapting images on small displays. Multimedia Syst 2003;9:353–64. [156] Lew D, Brinkman G, Ibanez E, Hodge B, King J. Western wind and solar
[126] Xu K, Ba J, Kiros R, Cho K, Courville A, Salakhudinov R, Zemel R, Bengio Y. integration study phase 2. Contract 2013;303:275–3000.
Show, attend and tell: Neural image caption generation with visual attention. [157] Hodge B-M, Orwig K, Milligan MR. Examining information entropy approaches
In: International conference on machine learning. 2015, p. 2048–57. as wind power forecasting performance metrics. National Renewable Energy
[127] Guo M-H, Xu T-X, Liu J-J, Liu Z-N, Jiang P-T, Mu T-J, Zhang S-H, Martin RR, Laboratory; 2012.
Cheng M-M, Hu S-M. Attention mechanisms in computer vision: A survey. [158] Bessa RJ, Miranda V, Botterud A, Wang J. ‘Good’ or ‘bad’ wind power forecasts:
Comput Vis Media 2022;1–38. A relative concept. Wind Energy 2011;14:625–36.

19
Y. Chu et al. Renewable and Sustainable Energy Reviews 198 (2024) 114391

[159] Mellit A, Kalogirou SA. Artificial intelligence techniques for photovoltaic [170] Kingma DP, Ba J. Adam: A method for stochastic optimization. 2014, arXiv
applications: A review. Prog Energy Combust Sci 2008;34:574–632. preprint arXiv:1412.6980.
[160] Xu B, Wang N, Chen T, Li M. Empirical evaluation of rectified activations in [171] Haykin S. Neural networks: a comprehensive foundation. 3rd ed.. Upper Saddle
convolutional network. 2015, arXiv preprint arXiv:1505.00853. River, NJ: Pearson Prentice Hall; 2008.
[161] Nair V, Hinton GE. Rectified linear units improve restricted Boltzmann [172] Lakhmi CJ, Martin NM. Fusion of neural networks, fuzzy systems and genetic
machines. In: Proc. int. conf. mach. learn.. ICML, 2010, p. 807–14. algorithms: industrial applications. CRC Press; 1998.
[162] Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks. In: Proc. [173] Fukushima K, Miyake S. Neocognitron: A new algorithm for pattern recog-
14th int. conf. artificial intell. stat. 2011, p. 315–23. nition tolerant of deformations and shifts in position. Pattern Recognit
[163] Werbos PJ. Beyond regression: new tools for prediction and analysis in the 1982;15:455–69.
behavioral sciences (Ph.D. thesis), Harvard University; 1974. [174] Al-Saffar AAM, Tao H, Talab MA. Review of deep convolution neural network
[164] Parker DB. Learning logic. Technical report TR-47, Center for Computational in image classification. In: Proc. int. conf. radar, antenna, microw. electron.
Research in Economics and Management Science, MIT; 1985. telecommun.. ICRAMET, 2017, p. 26–31.
[165] Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward [175] Deng L. The mnist database of handwritten digit images for machine learning
neural networks. In: Proc. int. conf. artificial intell. stat.. ICAIS, 2010, p. research [best of the web]. IEEE Signal Process Mag 2012;29:141–2.
249–56. [176] LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to
[166] He K, Zhang X, Ren S, Sun J. Delving deep into rectifiers: Surpassing human- document recognition. Proc IEEE 1998;86:2278–324.
level performance on imagenet classification. In: Proc. IEEE int. conf. comput. [177] Rawat W, Wang Z. Deep convolutional neural networks for image classification:
vis. 2015, p. 1026–34. A comprehensive review. Neural Comput 2017;29:2352–449.
[167] Pan SJ, Yang Q. A survey on transfer learning. IEEE Trans Knowl Data Eng [178] Ranzato M, Huang FJ, Boureau Y-L, LeCun Y. Unsupervised learning of invariant
2009;22:1345–59. feature hierarchies with applications to object recognition. In: Proc. IEEE
[168] Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in deep comput. vis. pattern recognit.. CVPR, 2007, p. 1–8.
neural networks? In: Proc. adv. neural inform. process. syst. 2014, p. 3320–8. [179] Lin M, Chen Q, Yan S. Network in network. 2013, arXiv preprint arXiv:
[169] Tan C, Sun F, Kong T, Zhang W, Yang C, Liu C. A survey on deep transfer 1312.4400.
learning. In: Artificial neural networks and machine learning–ICANN 2018: 27th
international conference on artificial neural networks, rhodes, Greece, October
(2018) 4-7, proceedings, part III 27. Springer; 2018, p. 270–9.

20