Knowledge-Based Systems 301 (2024) 112279

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

Deep learning methods for multi-horizon long-term forecasting of Harmful

Algal Blooms
Silvia Martín-Suazo a , Jesús Morón-López b , Stanislav Vakaruk a , Amit Karamchandani a ,
Juan Antonio Pascual Aguilar c , Alberto Mozo a ,∗, Sandra Gómez-Canaval a , Meritxell Vinyals d ,
Juan Manuel Ortiz c
a
E.T.S. de Ingeniería de Sistemas informáticos, Universidad Politécnica de Madrid, Calle Alan Turing s/n, Madrid, 28031, Spain
b
European Regional Centre for Ecohydrology of the Polish Academy of Sciences, 3 Tylna, Łódź, 90-364, Poland
c
IMDEA Water Institute, Avenida Punto Com, 2, Parque Científico Tecnológico de la Universidad de Alcalá, Alcalá de Henares, 28805, Spain
d
Université Fédérale de Toulouse, INRAE, MIAT, 24 Chem. de Borde Rouge, Castanet-Tolosan, 31326, France

ARTICLE INFO ABSTRACT

Dataset link: https://github.com/stanislavvaka The increasing occurrence of Harmful Algal Blooms (HABs) in water systems poses significant challenges
ruk/Chlorophyll_soft-sensor_machine_learning_ to ecological health, public safety, and economic stability globally. Deep Learning (DL) models, notably
models Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM), have been widely employed
Keywords: for HAB prediction. However, the emergence of state-of-the-art multi-horizon forecasting DL architectures
Deep learning such as Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) provides a novel
Chlorophyll solution for long-term HAB prediction. This study compares the performance of N-BEATS with LSTM and
Harmful algae bloom CNN models using high temporal granularity water quality data from As Conchas reservoir (NW Spain) to
Monitoring forecast chlorophyll-a (Chl-a) concentrations, a key indicator of HABs. The evaluation encompasses one-day
Time-series and one-week prediction horizons, aligning with World Health Organization (WHO) HAB alert criteria. Results
Forecasting
indicate that N-BEATS outperforms LSTM and CNN models for one-week predictions and when forecasting
multiple consecutive days within a week. Furthermore, augmenting input data with additional variables does
not significantly enhance predictive accuracy, challenging the assumption that complexity always improves
model performance. The study also explores the transferability of trained models across different monitoring
buoys within the same water body, emphasizing the adaptability and broad applicability of predictive models
in diverse aquatic environments. This research underscores the potential of N-BEATS as a valuable tool for
HAB prediction, particularly for longer-term forecasting.

1. Introduction 6: Ensure access to water and sanitation for all) [9], it is imperative to
intensify the monitoring of HABs and effectively coordinate climate and
In the Anthropocene era, in which human activity has had a signif- water-related actions.
icant impact on the planet’s climate and ecosystems, the relationship In this perspective, Artificial Intelligence (AI) is increasingly pro-
between societal needs and water resources has become very com- posed to modulate integrated water management [10]. Through a
plex [1]. This is evidenced by the widespread and record-breaking subset of algorithmic models based on techniques such as Machine
increase of Harmful Algal Blooms (HABs) in inland water bodies and Learning (ML) and Deep Learning (DL), AI is able to learn and predict
coastal areas around the world [2,3]. HABs are intensifying as a result outcomes through the observation of environmental changes. The use
of rising temperatures and high nutrient loads mainly from agriculture
of AI for HABs forecasting has been explored mainly with hyperspectral
and livestock runoff [4]. They, in turn, endanger the environment,
data collected by remote sensing [11]. Instruments aboard satellites and
animals and humans due to toxin production and dissolved oxygen
aircrafts are capable of measuring water reflectance signals resulting
depletion following bloom collapse [5,6]. Estimates point to an annual
from light absorption and scattering by algal cells. Numerous studies
cost of billions of dollars in public health and loss of ecosystem services
(e.g. fisheries, tourism and recreation) [7,8]. Therefore, if the United have applied ML and DL models to interpret these water-leaving re-
Nations Sustainable Development Goal on Water is to be achieved (Goal flectance signals and obtain information on the presence, abundance

∗ Corresponding author.
E-mail addresses: [email protected] (J. Morón-López), [email protected] (A. Mozo).

https://doi.org/10.1016/j.knosys.2024.112279
Received 22 March 2024; Received in revised form 7 July 2024; Accepted 24 July 2024
Available online 27 July 2024
0950-7051/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

and spatial distribution of HABs in surface waters [12–15]. However, and Long Short-Term Memory (LSTM) models, have also been inves-
although remote sensing offers wide spatial coverage, it only provides tigated within this context for their potential in managing temporal
daily or weekly data due to overflight limitations or interference caused dependencies in the data [23,38,39]. Notably, Barzegar et al. [23]
by cloud cover, complex water optics or weather conditions [16]. In ad- proposed a hybrid LSTM and Convolutional Neural Network (CNN)
dition, geographical resolution poses challenges for small water bodies, model, which demonstrated superior performance in predicting Chl-
as pixels may overlap with those of the Earth, making spatial estimation a. Traditional architectures, such as LSTM and CNN, have long been
difficult. This may lead to inaccurate predictions because the spatial the stalwarts in tasks requiring predictions over a fixed horizon. How-
and temporal distribution of HABs may show short-term variations ever, the advent of modern DL architectures specifically engineered
in surface waters. It is therefore necessary to combine the strengths for multi-horizon forecasting, represents a significant stride forward.
of remote sensing with complementary technologies operating at near These architectures can be trained in a sequence-to-sequence manner to
real-time scales to mitigate temporal uncertainty [16,17]. produce results across multiple prediction horizons in a single iteration,
High temporal granularity of water quality data is possible by de- eliminating the necessity to train separate models for each individual
ploying sensors directly in the water. There are sensors to measure the horizon. Not only do they offer enhanced computational efficiency, but
Chlorophyll-a (Chl-a), a pigment considered a proxy indicator of algal these models also bestow a greater degree of flexibility compared to
concentration in surface waters, as well as other physicochemical pa- their fixed-horizon counterparts. This allows end-users to select the
rameters related to HAB dynamics, such as pH, temperature, dissolved most suitable prediction horizon based on their unique requirements
oxygen, conductivity, and turbidity [18]. These sensors may be inte- and the tolerable bounds of prediction error. Recently, the domain
grated into buoys as Automatic High-Frequency Monitoring (AHFM) of interpretable DL algorithms has witnessed the inception of ‘‘Neural
systems to collect and transmit massive data in real time [19–21]. The Basis Expansion Analysis for Interpretable Time Series Forecasting’’
vast amount of data that water quality sensors are capable of collecting (N-BEATS), a recently multi-horizon prediction framework proposed
at scales of minutes or hours successfully meets AI requirements for by Oreshkin et al. [40]. This architecture consists of a succession of
complex water quality assessment [22]. Several studies have trained densely stacked blocks, each incorporating numerous fully connected
ML and DL models with long-term data to infer Chl-a concentration or layers interconnected through residual links. This model design offers
algae cell density [23–26]. These models may be particularly suitable unique advantages: (i) it evidences a broad adaptability across a diverse
to triggering various levels of alert before algae cells exceeds certain range of target domains without the need for specific modifications;
preset thresholds, thus serving as critical indicators of potential risks (ii) it is rapid to train; and (iii), importantly, it has the capacity
and guiding appropriate responses to HABs [24,27]. Typical alert levels to produce interpretable results. The N-BEATS architecture has been
for short-term responses proposed by the World Health Organization widely employed to address various traditional forecasting challenges,
(WHO) are based on factors such as Chl-a concentration, algal type and its applications extend to various sectors such as energy [41],
and abundance, toxin levels, and environmental conditions [28]. At healthcare [42,43] and telecommunications [44]. In light of these
vigilance level, routine monitoring continues with heightened aware- improvements and advances in other fields, it is relevant to evaluate
ness. The alert level 1 involves increased monitoring at least once the potential of the N-BEATS algorithm for HAB forecasting and alert
a week and advisories as algal cell counts rise. Increasing the fre- generation.
quency of monitoring is important considering that certain harmful The aim of this research is to improve HABs prediction by examining
algal species (especially cyanobacteria) can two- or threefold in a few the effectiveness of the N-BEATS algorithm in relation to the most com-
days in natural water bodies [28]. A warning arises with high cell monly used LSTM and CNN models. Specifically, this research delves
counts, visible blooms, and concerning toxin levels, leading to public into the application of DL techniques to a multi-horizon forecasting
advisories and potential restrictions. The alert level 2, triggered by problem, thus generating predictions with an hourly granularity and
severe conditions, demands immediate action, including water body extending up to a one-week horizon. The choice of these time scales is
closures and evacuation directives. In the recently established alert in line with WHO recommendations, which advocate an intensification
levels framework for HABs in recreational water bodies, the vigilance of sampling frequency as higher alert levels are reached [28]. A high
level is associated with Chl-a values ranging from 3–12 μg∕L. Alert degree of accuracy and breadth of predictions could transform water
1 and Alert 2 are activated when Chl-a levels fall within the range body management by reducing near-future uncertainty. In particular,
of 12–24 μg∕L, or if they surpass these values with a prevalence of the inherent versatility of the multi-horizon forecasting model allows
harmful algal species. For drinking water, the threshold values are experts to select from a variety of future projections, a feature that may
significantly lower. The vigilance level is triggered at concentration be highly beneficial in light of the typically observed deterioration of
of Chl-a of 1 μg∕L, and alert level 1 and 2 at concentrations of Chl-a prediction accuracy over extended horizons. A highlight of our research
between 1–12 μg∕L or higher. Although there are regional variations is the exploration of the ability to detect the initial alarm level of
in specific alert levels and corresponding actions, the imperative for HABs within a multi-horizon forecasting framework. The development
continuous vigilance, timely communication and collaborative efforts of the HABs alert level based on a forecasting of Chl-a concentration
remains constant. Therefore, the use of predictive models to aid in may provide experts with valuable early warnings, thus enabling more
sampling and triggering alerts is a pivotal strategy for implementing responsive approaches to water body monitoring. This advance in fu-
accurate and cost-effective early warning systems to guide coordinated ture alert prediction methodologies could greatly improve coordination
decision making in water bodies vulnerable to HABs. between mitigation strategies under a holistic approach.
Conventional ML algorithms (i.e. non-deep), such as Linear Regres- This study is drawn from a unique dataset obtained from the As
sion (LR), Classification And Regression Tree (CART) [25], Random Conchas reservoir (Galicia, NW Spain). This dataset, collected assidu-
Forest (RF) [25,29–31], and Support Vector Machine (SVM) [32,33], ously over a period of three years at 15-min intervals from two buoys
have proven to be adequate for predictions at very short timescales equipped with sensors, serves as an excellent context for examining the
(minutes) or to build low-cost soft-sensors. However, these models effectiveness of different DL techniques. Several DL techniques under
may fall short in dealing with complex nonlinear relationships and two distinct scenarios were evaluated: the first employs solely historical
short- to medium-term temporal dependencies (i.e. days to weeks) Chl-a measurements as input, while the second incorporates additional
present in water quality data. In this context, DL algorithms promise exogenous variables from the buoys, such as water temperature, pH,
to overcome the limitations of conventional ML models. For instance, a and conductivity. The objective of comparing these two scenarios was
basic DL architecture as Multi-Layer Perceptron (MLP) has been applied to assess whether augmenting the input data with these supplementary
in studies related to algae concentration prediction [34–37]. More variables significantly enhances the predictive accuracy of the models.
advanced DL models, including Recurrent Neural Networks (RNNs) The trend and periodicity of these additional variables (which correlate

2
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

with Chl-a as observed in our previous work [25]) are expected to 2.1. Time-series forecasting description
provide useful information about the future behavior of Chl-a and
enhance the predictive capability of the data. Additionally, we delved Time series represents a collection of data points sequentially gath-
into the concept of model transferability, assessing whether a model ered over uniform time intervals, characterized by their chronological
trained on data from one buoy may sustain its accuracy when applied to order. This temporal ordering is essential, as each point is interde-
another. The findings from this component of the study could provide pendent and contextually linked to its time of occurrence. Central
insight into the generalizability of the model across water bodies and to diverse scientific and practical fields, time-series analysis is a sta-
potentially pave the way for more efficient methods of training the tistical methodology employed to examine and predict data points
model, obviating the need for individual buoy-based training. gathered over a specific time span. Within this analysis, a fundamen-
In summary, this study represents a significant advancement in tal distinction is made between endogenous and exogenous variables.
the state-of-the-art for predicting algae concentration, thanks to the Endogenous variables are contingent upon the system under investi-
following noteworthy contributions: gation, while exogenous variables are influenced by external factors
that can affect the endogenous variables. In the context of time-series
• A comprehensive comparison is conducted between the recent forecasting, endogenous variables (𝑌𝑖 ) refer to the data points that are
DL technique N-BEATS and other well-established DL methods, predicted, whereas exogenous variables (𝑋𝑖 ) are not forecasted but
namely LSTM and CNN networks. This evaluation focuses on the can be supplied as input to the model to enhance the accuracy of the
prediction of HABs, which serves to enrich existing literature forecast.
while offering valuable insights into the relative performance of Within the realm of forecasting methodologies, several forecasting
these sophisticated methods. methods are typically employed. One-step-ahead forecasting consists
• A method is proposed to compare different forecasting methods in of predicting one time step ahead from the current moment using
the context of a multi-horizon forecasting problem. This research observed data and prior predictions to generate the forecast, which
challenges traditional methodologies in predictive modeling of can be utilized iteratively to produce indefinite future predictions (see
HABs by combining a granularity of one hour and a forecasting the top result One-step-ahead in Fig. 1). Nonetheless, this method may
horizon extending up to one week. not be optimal for long-term forecasting problems due to the potential
• A distinctive aspect of this research lies in the evaluation of fore- accumulation of errors from each intermediary prediction. Formally,
casting techniques with respect to their ability to predict the HABs we can define one-step-ahead forecasting as follows. Let 𝐼 represent the
distinct set of samples in a specific time-series dataset. Each sample 𝑖
alert level in a multi-horizon forecasting context (e.g., rather
within the set 𝐼 corresponds to a scalar input 𝑋𝑖,𝑡 ∈ R and a target
than relying on more standard metrics such as value prediction
𝑌𝑖,𝑡 ∈ R at every time step 𝑡 ∈ [0, 𝑇𝑖 ], where 𝑇𝑖 signifies the size of each
error). The introduction of these advanced techniques offers a
complete sample 𝑖 (i.e., the length of the time series). One-step-ahead
novel approach to effectively plan water body sampling, with
forecasting aims to predict the value of the target variable 𝑌𝑖,𝑡 + 1 in
the potential to significantly increase the effectiveness of HABs
the subsequent time step, given the prior predicted steps and historical
management and mitigation efforts.
data up to time 𝑡.
• This research represents a pioneering application of these ad-
vanced techniques to a meticulously collected dataset from As
Conchas reservoir. This unique dataset allows for a comparison 𝑌̂𝑖,𝑡+1 = 𝑓 (𝑌𝑡−𝐾∶𝑡 , 𝑋𝑖,𝑡−𝐾∶𝑡 ) (1)
not only based on models trained on historical Chl-a values,
In this equation, 𝐾 denotes the size of the previous temporal win-
but also on a set of additional exogenous variables acquired
dow, and 𝑓 represents a function that maps the inputs and history
from sensors on-board buoys. This comprehensive examination
to the predicted output at time 𝑡 + 1. The objective is to identify an
allows for a better understanding of the benefits and limitations
appropriate function 𝑓 that minimizes the error between the predicted
associated with integrating different sensors into AHFM systems.
output 𝑌̂𝑖,𝑡 + 1 and the ground truth value 𝑌𝑖,𝑡 + 1 on a validation set. The
• This study also provides an analysis of the transferability of
size of the past temporal window 𝐾 relies on the problem and must be
trained models between different buoys located in the same water
ascertained empirically.
body. The results of this analysis have substantial implications To create a forecasting sequence using one-step-ahead forecasting,
for the acceleration and efficiency of model training, leading to we start with an initial input 𝑋𝑖,0 and apply the model 𝑓 to predict the
new strategies for the deployment of these models in various output 𝑌𝑖,1 . Subsequently, we use the predicted output as the input for
environments. the next time step, that is, 𝑋𝑖,1 = 𝑌𝑖,1 , and iterate the process to generate
a sequence of predicted outputs 𝑌𝑖,2 , 𝑌𝑖,3 , … , 𝑌𝑖,𝑇𝑖 . As evident, errors
This article is structured as follows. Firstly, a theoretical background
in each intermediary prediction can propagate to later predictions,
is provided in Section 2 in order to facilitate the knowledge to com-
causing a compounding effect on the overall forecast accuracy.
prehend the solution proposed in this article. Secondly, the proposed
In opposition, fixed-horizon forecasting is tailored to predict a
solution is presented in Section 3, offering a detailed description of the
specific future time step with optimized accuracy, thereby avoiding
procedures undertaken to prepare the experiments performed. In Sec-
error accumulation (see the central result Fixed-horizon in Fig. 1).
tion 4, the results of the proposed methods are exhaustively evaluated However, this technique is restricted to predicting only the particular
and discussed. Finally, the study’s main findings and conclusions are step and cannot be employed for earlier or later predictions. Fixed-
detailed in Section 5. horizon forecasting focuses on predicting the target variable 𝑌𝑖,𝑇 at a
specific future time 𝑇 , given the input variables 𝑋𝑖,𝑇 and the history
2. Theoretical framework of the time series up to time 𝑡, i.e., (𝑋𝑖,0 , 𝑌𝑖,0 ), … , (𝑋𝑖,𝑡 , 𝑌𝑖,𝑡 ). This can be
formalized as follows:
In this section, we will delve into the concept of time-series forecast-
ing and explore the various methods employed in this domain. Addi-
𝑌̂𝑖,𝑇 = 𝑓 (𝑌𝑡−𝐾∶𝑡 , 𝑋𝑖,𝑡−𝐾∶𝑡 ) (2)
tionally, we provide an in-depth explanation of the theoretical aspects
related to several ML/DL architectures widely used in the forecasting In this formula, 𝑇 represents the time at which the prediction
field, and we examine the fundamental principles and mechanisms occurs, 𝑌𝑡−𝐾∶𝑡 denotes the past values of the time series up to time
underlying these architectures and highlight their respective strengths 𝑡, and 𝑋𝑖,𝑡−𝐾∶𝑡 corresponds to the input variables associated with the
and applications in the context of forecasting tasks. time series over an identical time window. The function 𝑓 maps the

3
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 1. Graphical representation of the input–output structure of a prediction model according to the type of horizon: One-step-ahead, Fixed-horizon, and Multi-horizon.

historical values and inputs to the predicted value 𝑌̂𝑖,𝑇 at time 𝑇 . The that the N-BEATS architecture demonstrates better generalization abili-
goal is to find an appropriate function 𝑓 that minimizes the prediction ties than other DL models when trained on a specific source time-series
error between the predicted value 𝑌̂𝑖,𝑇 and the ground truth value 𝑌 𝑖, 𝑇 dataset and applied to a different target dataset [40,42–44,47].
on a validation set. In order to ensure a fair and meaningful comparison between N-
Multi-horizon forecasting offers an alternative approach that en- BEATS and other well-established machine learning and deep learning
compasses predicting a series of future steps (see the bottom result models, we conducted a thorough review of recent literature on time-
Multi-horizon in Fig. 1). This method strives to provide flexibility series forecasting. The primary objective of this review was to identify
in the forecasting process and minimize cumulative errors, although the most relevant models for comparison, taking into consideration
it may not achieve the same level of accuracy as the fixed-horizon their performance and applicability in the context of time-series fore-
method for predicting a singular point in time [45]. The multi-horizon casting. Among the wide range of architectures available, the Long
forecasting technique involves predicting the target variable for a series Short-Term Memory (LSTM) neural network architecture was selected
due to its exceptional performance in capturing temporal dependencies
of future time steps, i.e., 𝑌𝑖,𝑡 + 1 , 𝑌𝑖,𝑡 + 2 , … , 𝑌𝑖,𝑇 . This can be formalized
within univariate and multivariate time-series forecasting, as high-
as follows:
lighted in several studies [48–50]. LSTMs are particularly suitable
𝑌̂𝑖,𝑡 + 1∶𝑡 + ℎ = 𝑓ℎ (𝑌𝑡−𝐾∶𝑡 , 𝑋𝑖,𝑡−𝐾∶𝑡 ) (3) for time-series forecasting tasks that rely on long-term dependencies.
They excel at modeling sequential data by effectively capturing trends,
In the expression above, ℎ denotes the forecast horizon, 𝑌̂𝑖,𝑡∶𝑡 + ℎ rep- seasonality, and irregular patterns. Unlike traditional recurrent neural
resents the predicted sequence of the ℎ-step-ahead forecast with respect networks, LSTMs are capable of handling input sequences of varying
to the time step 𝑡, 𝐾 indicates the size of the past temporal window, and lengths and are resistant to the issues of vanishing and exploding gra-
𝑓 represents a function that maps the inputs and history up to time 𝑡 to dients. Their ability to retain information over extended periods makes
the predicted sequence 𝑌̂𝑖,𝑡 + 1∶𝑡 + ℎ . The objective is to find a suitable them well-suited for forecasting tasks that require the consideration of
function 𝑓ℎ for each horizon ℎ that minimizes the error between the extensive historical context.
predicted output 𝑌̂𝑖,𝑡∶𝑡 + ℎ and the ground truth value 𝑌𝑖,𝑡∶𝑡 + ℎ on a Convolutional Neural Networks (CNNs), originally designed for im-
validation set, typically using a metric that assesses the overall accuracy age processing, have also gained attention in the domain of time-series
of the predictions, such as the mean squared error or the mean absolute forecasting [47,48]. By leveraging the concept of convolutional oper-
error. However, in the case of using DL, a single model can be trained ations, CNNs can effectively extract local features and patterns from
to predict multiple horizons concurrently, an approach that is referred sequential data. This characteristic makes CNNs suitable for capturing
as Multi-output Multi-input (MIMO) forecasting [46]. In this approach, temporal dependencies within time-series data, such as trends and
the model is trained to predict the target variable for multiple future recurring patterns. Several studies have highlighted the effectiveness of
time steps, i.e., 𝑌𝑖,𝑡 + 1 , 𝑌𝑖,𝑡 + 2 , … , 𝑌𝑖,𝑇 , simultaneously. This technique CNNs in time-series forecasting tasks [47,48]. By employing convolu-
allows for increased flexibility and accuracy in long-term forecasting, tions across different dimensions of the input data, CNNs can effectively
as it can capture complex nonlinear relationships between the input capture single variable and cross-variable temporal dependencies. This
and output variables, and errors do not accumulate over time. ability is particularly advantageous in time-series analysis, as it enables
the network to learn complex relationships and interactions between
different variables, thereby potentially enhancing overall predictive
2.2. Selection of methods for time-series forecasting
performance. Additionally, CNNs are computationally efficient due to
shared weights and parallel processing, enabling them to handle large-
This study aims to assess the performance of the N-BEATS ar-
scale time-series datasets efficiently and making them well-suited for
chitecture as representative of state-of-the-art sequence-to-sequence
real-time applications.
time-series forecasting method, in comparison to traditional ML/DL This study also combined CNNs and LSTM networks to build a
architectures for the given problem. The selection of N-BEATS as hybrid model. The integration of CNNs and LSTMs aims to leverage the
the DL architecture was motivated by its exceptional performance in complementary strengths of both architectures and enhance the predic-
well-established forecasting competitions [40]. As commented above, tive capabilities of the overall model in time-series forecasting tasks.
N-BEATS is specifically designed for time-series forecasting and has The hybrid model combining CNNs and LSTMs offers the potential to
demonstrated remarkable results with univariate time-series data. The capture both local and global temporal dependencies within time-series
N-BEATS architecture offers several advantageous characteristics. It data. By employing CNNs for initial feature extraction and leveraging
can be directly applied to various problem types without requiring the ability of LSTMs to capture long-term dependencies, the hybrid
extensive feature engineering. Furthermore, it exhibits faster training model can effectively capture both short-term patterns and long-term
and greater scalability compared to other DL architectures like LSTM. trends in the data. Several studies have demonstrated the effectiveness
Additionally, it possesses model interpretability capabilities that are of this hybrid approach in improving the predictive performance of
lacking in other DL architectures [40]. Moreover, studies have shown time-series forecasting models [51–53].

4
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

2.3. Overview of selected methods for time-series forecasting

In this subsection, we present a comprehensive analysis of the the-

oretical aspects of several Machine Learning (ML) and Deep Learning
(DL) architectures employed for forecasting tasks.

2.3.1. Fully connected neural network

Fully Connected Neural Networks (FCNNs), also referred as multi-
layer perceptrons, are a fundamental form of feed-forward neural net-
works extensively utilized in DL. These FCNNs consist of interconnected
neurons organized into layers, with each neuron in a given layer
connected to all neurons in the subsequent layer, resulting in a dense
network structure [47]. The layers within an FCNN can be categorized
into three types: the input layer, hidden layers, and output layer. Fig. 2. Architecture of a long short-term memory cell, showing the interactions
The choice of the number of neurons in the hidden layers is crucial between the input 𝑋𝑡 , the previous hidden state 𝐻𝑡−1 , and the previous cell state
and involves striking a balance between model complexity and gen- 𝐶𝑡−1 . 𝜎 denotes the sigmoid function, while tanh corresponds to the hyperbolic tangent
function. The symbols × and + represent element-wise multiplication and addition,
eralization capability. A higher number of neurons allows the FCNN
respectively. 𝐻𝑡 denotes the current hidden state, and 𝐶𝑡 denotes the current cell state,
to learn complex patterns by capturing a larger number of features, both of which are updated based on the inputs and the outputs of the forget, input,
but too many can cause overfitting, leading to poor performance on and output gates. The updated values 𝐻𝑡 and 𝐶𝑡 are then passed to the next LSTM cell
unseen data. On the other hand, a lower number of neurons may limit in the sequence.
the expressive power of the FCNN, hindering its ability to capture com-
plex relationships in the data. Proper initialization and regularization
techniques (e.g., Dropout, L2, etc.) are often employed to prevent issues The forget gate determines the extent to which the previous memory
such as vanishing or exploding gradients, ensuring stable and effective cell state, denoted as 𝐶𝑡−1 , should be forgotten or preserved. The forget
learning within the hidden layers. The output layer of an FCNN receives gate activation 𝐹𝑡 is computed as:
the results of the consecutive non-linear data transformations achieved
through the hidden layers and learns how to map these non-linear 𝐹𝑡 = 𝜎(𝑊𝑥𝑓 ⋅ 𝑋𝑡 + 𝑊ℎ𝑓 ⋅ 𝐻𝑡−1 + 𝑏𝑓 ) (5)
feature representation, or embedding, to the desired target [47]. These where 𝑊𝑥𝑓 , 𝑊ℎ𝑓 , and 𝑏𝑓 are the weight matrix and bias term associated
output neurons generate the final output of the FCNN, which could be with the forget gate.
a prediction, classification, or any other relevant outcome, based on the
The memory cell 𝐶𝑡 stores the information over time and is updated
specific task the model is designed to perform.
using the input and forget gates. The updated memory cell state is
One noteworthy limitation of FCNNs is their inability to effectively
computed as:
leverage the inherent data structures commonly present in forecasting
applications [47]. Moreover, their fixed number of inputs and outputs 𝐶𝑡 = 𝐹𝑡 ⋅ 𝐶𝑡−1 + 𝐼𝑡 ⋅ tanh(𝑊𝑥𝑐 ⋅ 𝑋𝑡 + 𝑊ℎ𝑐 ⋅ 𝐻𝑡−1 + 𝑏𝑐 ) (6)
makes them unsuitable for problems that involve varying input and
output sizes, as typically encountered in forecasting tasks [47]. In where 𝑊𝑥𝑐 , 𝑊ℎ𝑐 , and 𝑏𝑐 are the weight matrix and bias term associated
order to overcome these limitations, specialized architectures have with the memory cell.
been developed [48] that incorporate FCNNs as fundamental building The output gate determines the extent to which the current memory
components in the form of Fully Connected (FC) layers. In the following cell state 𝐶𝑡 should be exposed to the next hidden state 𝐻𝑡 . The output
discussion, we explore these advanced architectures that effectively gate activation 𝑂𝑡 is computed as:
address the challenges mentioned above.
𝑂𝑡 = 𝜎(𝑊𝑥𝑜 ⋅ 𝑋𝑡 + 𝑊ℎ𝑜 ⋅ 𝐻𝑡−1 + 𝑏𝑜 ) (7)

2.3.2. Long-short term memory where 𝑊𝑥𝑜 , 𝑊ℎ𝑜 , and 𝑏𝑜 are the weight matrix and bias term associated
Long Short-Term Memory (LSTM) [54] networks were introduced with the output gate.
to address the vanishing gradient problem present in RNNs. This issue The hidden state 𝐻𝑡 is the output of the LSTM at time step 𝑡.
arises from the exponential decay of gradients over time, hindering Mathematically, the hidden state 𝐻𝑡 is computed as:
learning over lengthy sequences [47]. The LSTM architecture enables
𝐻𝑡 = 𝑂𝑡 ⋅ tanh(𝐶𝑡 ) (8)
the model to capture and preserve long-term dependencies, making
them particularly effective for tasks that involve complex patterns LSTMs effectively address the problem of gradient vanishing, pre-
across long periods of time. To that end, as depicted in Fig. 2, the venting gradients from dissipating as they traverse the network [48,55].
internal structure of the LSTM introduces specialized memory cells This is achieved through the aforementioned gate-based mechanism,
and gating mechanisms that allow it to selectively remember or forget which allows the cell state to be preserved and updated as neces-
information at different time steps. These gating mechanisms consist of sary. LSTMs are commonly trained using a variation of the classical
three main gates: the input gate, the forget gate, and the output gate. backpropagation algorithm known as Backpropagation Through Time
The input gate controls the incorporation of new information into (BPTT). While similar to traditional backpropagation, BPTT accounts
the memory cell. It applies a sigmoid activation function to the linear for the LSTM’s ability to predict sequences comprising multiple time
combination of the current input, denoted as 𝑋𝑡 , and the previous hid- steps rather than a single value. To accomplish this, BPTT unfolds
den state, denoted as 𝐻𝑡−1 . Mathematically, the input gate activation the LSTM network over time and computes the gradients of the cell
𝐼𝑡 is defined as: state and gates at each time step. Consequently, the loss gradient with
respect to the cell state at time 𝑡 is backpropagated through time to
𝐼𝑡 = 𝜎(𝑊𝑥𝑖 ⋅ 𝑋𝑡 + 𝑊ℎ𝑖 ⋅ 𝐻𝑡−1 + 𝑏𝑖 ) (4)
the cell state at time 𝑡−1 . Notably, the network parameters are shared
where 𝑊𝑥𝑖 , 𝑊ℎ𝑖 , and 𝑏𝑖 are the weight matrix and bias term associated across all time steps. Once the gradients are computed, they propagate
with the input gate. through the unrolled network [48].

5
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 3. CNN network architecture and layers.

2.3.3. Convolutional neural networks 2.3.4. N-BEATS

Convolutional neural networks (CNNs) have emerged as a powerful The Neural Basis Expansion Analysis for Interpretable Time Series
and increasingly popular DL architecture for forecasting tasks. The Forecasting (N-BEATS) architecture, introduced by Oreshkin et al. [40],
inception of CNNs can be traced back to the year 1989 when [56] presents a novel approach to time-series forecasting that has shown
initially proposed this concept, drawing inspiration from Fukushima’s exceptional performance in renowned forecasting competitions and
Neocognitron model introduced in 1980 [57]. Since then, CNNs have benchmarks. N-BEATS is a DL model explicitly designed for time-series
been widely adopted for image classification, object detection, and forecasting, exhibiting impressive results particularly with univariate
other computer vision tasks. Although initially developed for image time-series data. One aspect that sets N-BEATS apart from other DL
analysis, CNNs have proven their effectiveness in a variety of fields architectures is its ability to provide interpretability of the predic-
where data possess a known ordinal structure [47], such as time-series tions. This is achieved by explicitly breaking down the forecasting task
prediction. into interpretable components, including trend and seasonality. This
CNNs consist of multiple layers that operate hierarchically to cap- explicit breakdown allows analysts and practitioners to gain insights
ture temporal patterns and make predictions [47]. As depicted in Fig. 3, and understanding into the underlying factors driving the forecasts.
these layers include convolutional layers, pooling layers, and FC layers, By offering this level of transparency and interpretability, N-BEATS
each serving a specific role in the network [47]. Convolutional layers surpasses traditional statistical techniques, making it a preferred choice
are responsible for extracting relevant features from the input time- in practical scenarios where interpretable forecasting is crucial [40].
series data by applying convolution operations over sliding windows The basic block of the N-BEATS architecture consists of two com-
of the input sequence. This allows the network to automatically learn ponents: FCNNs and basis layers. These blocks are arranged vertically
in an arbitrary number of stacks to form the complete architecture.
meaningful representations at different temporal scales. Pooling layers
In the first block, represented as 𝑥𝑙 , the input 𝑋𝑙 corresponds to a
play a crucial role in downsampling the extracted features while retain-
historical window of observations with a specific length 𝑙, ending with
ing their essential characteristics and enhancing the network’s ability
the most recent measurement. Originally, 𝑙 is chosen as a multiple of
to generalize and capture important temporal patterns. The output of
the forecast horizon denoted as 𝐻. Each block takes an input, denoted
the convolutional and pooling layers is then fed into FC layers, which
as 𝑥𝑙 , and produces two outputs: a forward estimate 𝑦̂𝑙 (forecast) and
integrate the learned representations and make predictions based on
a backward estimate 𝑥̂ 𝑙 (backcast) of 𝑥𝑙 . 𝑦̂𝑙 represents the block’s
the extracted features.
predicted values for the future time steps, while the 𝑥̂ 𝑙 represents the
One of the key advantages of CNNs for forecasting is their ability to
best approximation of the input 𝑥𝑙 based on the constraints imposed
automatically learn hierarchical representations from sequential data.
on the block’s functional space. In subsequent blocks, the input 𝑥𝑙 is
By exploiting the hierarchical nature of convolutional layers, CNNs
obtained as the residual output of the previous block, specifically the
can initially capture local temporal correlations, progressively learning
backcast generated by the preceding block.
more abstract and high-level features in deeper layers. This hierarchical The aggregation of forecasts in the N-BEATS architecture follows
learning allows CNNs to extract relevant information across multiple a hierarchical doubly residual stacking principle based on the topol-
levels of abstraction in the temporal dimension, thereby enhancing ogy originally proposed by other well-known architectures such as
their predictive capabilities. Furthermore, the shared weights of CNNs DenseNet [59]. This principle involves two residual branches: one for
enable the network to learn spatially invariant features, facilitating the the backcast estimate and another for the forecast branch within each
application of the same learned feature across different regions of the block. The backcast branch allows for sequential analysis of the input
input sequence. Additionally, the local connectivity inherent to CNNs signal, progressively removing approximated components at each block
ensures that each neuron is responsible for a specific receptive field, to assist downstream blocks in their forecasting tasks. This structural
effectively capturing local dependencies within the input data. This design facilitates gradient backpropagation and ensures smooth infor-
characteristic makes CNNs well-suited for analyzing sequential infor- mation flow throughout the network. The final forecast is obtained by
mation, as neighboring time steps are more likely to exhibit stronger summing all the partial forecasts generated by the blocks.
relationships compared to those that are further apart. Each block is composed of multiple FC layers, with each layer
The application of CNNs in time-series forecasting requires cer- being followed by a Rectified Linear Unit (ReLU) activation function.
tain adaptations to accommodate the temporal nature of the data. These stacked FC layers are responsible for predicting the forward
Unlike 2D CNNs used on two-dimensional grids, 1D CNNs process one- expansion coefficients (𝜃𝑙𝑓 ) and backward expansion coefficients (𝜃𝑙𝑏 ).
dimensional sequences [48]. Convolutional and pooling operations in Subsequently, the basis layers, denoted as 𝑔𝑙𝑏 and 𝑔𝑙𝑓 , perform a linear
1D CNNs are applied along the temporal axis, rather than across spatial transformation based on an element-wise product between the expan-
dimensions as in 2D CNNs. Moreover, in contrast to conventional sion coefficients of each branch and the corresponding weights of the
ML/DL, CNNs can efficiently handle multivariate data by treating each basis layers to generate the backcast and forecast outputs of the block.
variable as a separate channel, with convolutions applied indepen- A complete diagram of the architecture and its different components is
dently to each channel [58]. FC layers at the end of the CNN then presented in Fig. 4.
extract dependencies between these channels, enabling comprehensive The flexibility of the N-BEATS architecture allows for the choice of
multivariate time-series analysis. learnable basis layers or predefined functional forms, depending on the

6
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 5. Summary of the data acquisition, preprocessing and deep learning models
training and validation workflow.

Fig. 4. Diagram of the N-BEATS architecture showing the internal structure of a stack, of one hour and a forecast horizon of one week into the future. For the
including the residual connections between blocks and the layers within a block. first time, the N-BEATS architecture is introduced for Chl-a forecasting
and its performance is evaluated against traditional DL architectures
such as CNN and LSTM. Previous research has demonstrated the su-
specific problem and desired inductive biases for the task. By imposing perior performance of N-BEATS in multi-horizon forecasting across
constraints on the weights of the basis layers, the architecture can be various domains [60]. Furthermore, this study explores an innova-
adjusted to reflect problem-specific constraints in the generated outputs tive application that utilizes a multi-horizon Chl-a forecasting model
and enable subsequent analysis of the predictions by decomposing them to identify algae proliferation alert levels, thereby transforming the
into trend and seasonality components, as well as other relevant factors, problem into a multi-horizon classification task.
and consequently allowing for interpretation of the decisions made by In this section, we will outline the steps taken from data acquisition
the model. However, a generic implementation of N-BEATS commonly to the training and evaluation of multi-horizon forecasting models for
employs learnable basis layers, as they offer greater flexibility and Chl-a values in a freshwater body. Section 3.1 will provide details on
adaptability to various forecasting tasks.
the environment and sensors used to collect data from the freshwater
The N-BEATS architecture possesses numerous advantageous char-
body over a period of three years. Section 3.2 will describe the steps
acteristics. Firstly, it can be directly applied to diverse problem types
followed to process the captured sensor data. Section 3.3 will explain
without requiring extensive feature engineering, thus saving consider-
able time and effort. Moreover, N-BEATS outperforms other DL archi- how the models were trained, selected, and evaluated. Finally, Sec-
tectures such as LSTM in terms of training efficiency and scalability, tion 3.4 will specify the technology and dataset used for conducting
which is especially advantageous when handling extensive datasets. the experiments performed in this study. A summary of the workflow
Additionally, the interpretability aspect of N-BEATS is of significant of this process can be appreciated in Fig. 5.
value, allowing users to gain insights into the model’s decision-making
process. Furthermore, studies have demonstrated the superior gener- 3.1. Data acquisition
alization capabilities of the N-BEATS architecture when trained on
a specific source time-series dataset and applied to a distinct target
dataset [40,42–44,47]. These findings underscore the robustness and Located in the Miño-Sil River Basin District of Galicia, NW Spain,
adaptability of the N-BEATS model, further reinforcing its suitability the As Conchas reservoir forms an integral part of the ‘‘Baixa Limia-
for time-series forecasting tasks. Serra do Xurés’’ Natural Park. This eutrophic freshwater body can store
While originally conceived as a univariate model, a novel redesign up to 80 Hm3 of water, displaying pronounced depth variations, from
of N-BEATS that empowers it to handle multivariate series was pre- shallow edges to a central depth of 32 m at full capacity. According
viously proposed by [60]. This redesign allows for the consideration to the Miño-Sil Hydrographic Confederation, the depth of the photic
of variables based on their relationship with independent variables, zone of the reservoir measured with the Secchi disk (indicative of water
whether they are endogenous or exogenous. To that end, a concate- clarity) ranges between 2 and 6 m from May to October over a three-
nation layer that merges multiple inputs into a one-dimensional input year observation period. Two centrally anchored EM1250 buoys from
layer is introduced. By adopting this approach, the model is optimized Xylem Analytics, designated as the beach buoy and the dam buoy, fa-
solely to capture the dynamics of endogenous variables, with exogenous cilitate reservoir monitoring. Equipped with YSI multiparametric EXO3
variables serving as support for model learning rather than being the probes, these buoys housed sensors that measured Chl-a using a total
forecasting target. This design choice ensures that the model’s capa-
algae fluorescence sensor, pH, water temperature, and EC, as detailed
bilities are entirely dedicated to modeling the endogenous dynamics,
by Morón-López et al. [19]. These probes and sensors were subjected
without being encumbered by the entire multivariate context.
to regular maintenance and calibration according to manufacturer’s
3. Multi-horizon prediction framework setup instructions, ensuring reliable data collection. Solar panels provided
the necessary power for this autonomous system, with battery levels
The objective of this study to develop the first multi-horizon fore- serving as an indicator for daylight hours. The probes, positioned at a
casting model for Chl-a levels in a freshwater body, with a granularity standard depth of approximately 1 m, were set to log data every 15 min.

7
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Table 1 Table 3
Statistical overview of the dataset characteristics for the beach buoy, which comprises Total count of data rows associated with each dataset segment (training, weight
approximately 28,000 samples. validation, hyperparameter validation, and testing) following the data partitioning
Statistic Temperature pH Conductivity Chlorophyll process for both beach and dam buoys. The final column specifies the cumulative total
(◦ C) (μS/cm) (μg/L) of rows for each respective buoy.
Dataset Training Weights Hyperparameter Testing Total
Average 15.10 7.74 66.85 8.86
validation validation
Standard deviation 6.07 0.87 15.66 7.67
Minimal value 5.38 6.56 1.03 0.00 Beach buoy 15 640 1912 1646 8398 27 596
Percentile 25% 9.38 7.11 55.34 3.98 Dam buoy 14 989 1912 1646 8397 26 944
Percentile 50% 14.36 7.44 70.74 6.56
Percentile 75% 21.03 8.11 77.22 11.22
Maximal value 27.75 10.26 99.20 78.65
period. Therefore, it was decided to resample the dataset at a coarser
granularity, to hourly intervals, using the statistical function of median.
Table 2
Statistical overview of the dataset characteristics for the dam buoy, which comprises Resampling the data helps expediting the training of deep learning
approximately 27,000 samples. models without sacrificing important information. Additionally, the
Statistic Temperature pH Conductivity Chlorophyll median function serves to mitigate the number of outliers that may
(◦ C) (μS/cm) (μg/L) arise due to measurement errors over an hour-long period.
Average 15.11 7.43 67.69 9.09
Generation of time series. Diverse window sizes have been selected
Standard deviation 5.94 0.94 9.04 8.30
Minimal value 6.15 4.93 4.35 0.00 according to the number of time steps used to attain the forecasts (input
Percentile 25% 9.41 6.88 62.05 3.59 window size) and the future time steps that the models will forecast
Percentile 50% 14.15 7.16 67.82 7.07 (forecast horizon size). Input window sizes were studied in the set {12,
Percentile 75% 21.10 7.63 73.06 11.95 24, 72, 168, 336} whereas forecast horizon sizes were studied in the set
Maximal value 28.48 10.30 97.57 71.66
{24, 168}. The alert level 1 involves increased monitoring at least once
a week, then the one-day or one-week ahead forecasting would aid the
planning of the water-body monitoring. It is important to note that the
3.2. Data preprocessing maximum input window size was set to be at least twice the maximum
forecasting horizon size, and the minimum window size was set to be
DL models require appropriate data formatting to ensure effective at least half of the minimum forecasting horizon size. Additionally,
training for the specific problem at hand. For the multi-horizon, it intermediate window sizes were defined to assess the impact of window
was crucial to adapt the data to a time series format and split it size on the deep learning models.
into separate training, testing, and validation datasets. Additionally, Time series cleaning . To ensure the models are not affected by noise
DL models can benefit from outlier data cleaning and conditioning and can effectively learn, all rows in the time window and forecasting
procedures. In this study, the dataset used was collected remotely from horizon that contained any invalid data points were excluded. These
two sensor-equipped buoys. Due to occasional environmental effects, invalid data points can arise from occasional sensor errors or missing
there were instances of data loss in the sensor-to-server communication, data due to communication interference. In this particular case, ap-
and occasional erroneous measurements were recorded by the sensors. proximately 4% of the data was identified as invalid and subsequently
As a result, prior to training the DL models for Chl-a level forecasting, discarded.
it was necessary to perform data cleaning and correction processes.
The data preprocessing process employed in this study consists of Dataset splitting . To mitigate overfitting during model training, the
the following steps: (i) feature selection for the training process; (ii) original dataset was split into four distinct subsets, while maintaining
resampling the data from 15 min to 1 h intervals; (iii) constructing the the sequential order of the data for each buoy. These subsets were
time series; (iv) cleaning and removing invalid data; (v) splitting the designated as follows: training, weights validation, hyperparameter
data into training, testing, and validation sets; (vi) normalizing the data validation, and testing. To ensure models’ generalizability, the data
and removing seasonality. In what follows, we describe in detail each from the last year was reserved exclusively for the testing dataset. Prior
of preprocessing steps. to that, 10% of the data before the last year was allocated for weights
validation, and another 10% was set aside for hyperparameter valida-
Feature selection. The Chl-a variable acts as the target forecast vari- tion. The remaining data comprised the training dataset. Fig. 6 presents
able, and following standard forecasting systems practice, its historical a detailed depiction of the temporal evolution of Chl-a concentrations
observed values are utilized as inputs for the DL models. In this context, within the testing partition, encompassing data from both the beach
Chl-a is regarded as the endogenous variable. Furthermore, exogenous and dam buoy locations.
variables were incorporated to enhance the quality of the trained mod- The model was initially trained using the training dataset, and the
els. More specifically, temperature, pH, and conductivity were selected loss function optimization was monitored using the validation dataset.
as exogenous variables due to their high correlation with Chl-a, as Subsequently, the model was evaluated using the hyperparameter vali-
reported in the study by Mozo et al. [25] using the same dataset. These dation dataset to determine the optimal set of hyperparameters. Table 3
four features were employed in conjunction with the Chl-a endogenous provides a breakdown of the number of data rows in each of the
variable to improve the models obtained solely using the endogenous resulting datasets obtained from the splitting process.
variable. However, when training solely with exogenous variables,
Data conditioning . The data was preprocessed using standard scaling
there was no observed improvement in the results compared to the
and seasonality removal techniques. Standard scaling is a method that
other variable combinations. Hence, in our experiments, we considered
normalizes the data’s mean to zero and standard deviation to one. Sea-
Chl-a alone or Chl-a with exogenous variables (often referenced as Chl-
sonality removal involves eliminating patterns that periodically repeat
a+Exo). For a more detailed analysis of the statistical nature of the
to mitigate their influence. Preconditioning the dataset in this manner
endogenous and exogenous variables, please refer to Tables 1 and 2,
facilitates the training of DL models and prioritizes the learning of more
corresponding to the beach and dam datasets, respectively.
relevant patterns.
Data resampling. The buoys measured and transmitted data to a server In this study, a total of eight preprocessed datasets were generated
every 15 min. These data exhibited low variation between sequential from two distinct buoy sources: the beach buoy and the dam buoy.
samples due to minimal changes in the water masses within a 15-min These datasets were meticulously prepared to facilitate the experiments

8
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 6. Temporal progression of Chl-a concentrations within the testing dataset, incorporating data from both the beach and dam buoy sites. The horizontal axis denotes time in
hours, while the vertical axis represents Chl-a concentration in μg∕l.

conducted as part of this research. For each buoy, four datasets were model. In this CNN-LSTM hybrid model, an additional hyperparameter
processed, comprising training, validation, hyperparameter validation, called skipping connections is considered. It controls whether the output
and testing datasets. From these eight datasets, a careful selection of the first 1D-CNN network is concatenated with the input of the
was made, incorporating all four datasets from the beach buoy and subsequent 1D-CNN network. The other hyperparameters in Table 4
exclusively the testing dataset from the dam buoy. are standard for each architecture, while the hyperparameters in the
The primary objective of this study was to showcase the predic- ‘‘Common’’ row are shared among all architectures. All sets of hyper-
tive capacity of modern DL models in forecasting chlorophyll levels parameters were sampled following a uniform distribution to cover the
up to one week ahead in a body of water, while also evaluating breadth of the ranges more effectively.
their generalization capabilities. Consequently, the models were trained In these experiments, various combinations of the architectures
and rigorously evaluated using data exclusively from the beach buoy, listed above, input datasets, input window sizes, and forecast horizon
followed by thorough testing on the dam buoy. This approach was sizes were evaluated. Two types of input datasets were used: Chl-a only
adopted to demonstrate the models’ remarkable ability to generalize and Chl-a with exogenous variables. Five different input window sizes
effectively across different locations within the body of water. were explored: 12, 24, 72, 168, and 336 h. Two forecast horizon sizes
were examined: 24 and 168 h. For each model and scenario, up to
3.3. Deep learning models training and evaluation 30 sets of random hyperparameters were tested, resulting in a total of
approximately 2400 trained models. Table 5 provides a summary of the
In supervised problems, DL models are typically trained iteratively hyperparameter configurations explored in this study. Subsequently,
by optimizing their weights using observed labels from a given train- the best-performing set of hyperparameters, based on the performance
ing dataset, while simultaneously monitoring convergence through a on the hyperparameter validation dataset, was selected for each model,
separate validation dataset. Prior to training, it is crucial to define a resulting in 80 scenarios for comparison on the test dataset.
set of hyperparameters. However, since there is no universally optimal Multi-horizon forecasting models are commonly assessed using met-
set of hyperparameters that applies to all problems, it is necessary to rics for measuring the accuracy of regression models, such as Mean
explore different combinations and identify the most suitable ones for Absolute Error (MAE) or Mean Squared Error (MSE), which provide
each specific model and scenario. To address this challenge, a random a measure of the average deviation between the model’s predictions
search heuristic is employed to generate diverse sets of hyperparameter and the expected values. While MAE considers the absolute difference,
values randomly sampled from predefined hyperparameter grids for MSE additionally penalizes larger errors more severely, making it par-
each model or scenario. Each of these models is then trained on the ticularly suitable for problems where significant deviations have higher
training dataset and evaluated on a separated validation dataset. Ulti- importance. In this study, both metrics were employed to evaluate the
mately, the best-performing model is selected based on its performance accuracy of the regression models. Additionally, the MSE was utilized
on the validation dataset. as the loss function during the training phase to optimize the model
The random search phase explored various hyperparameters, in weights, thereby placing a greater penalty on larger errors.
selected ranges, and the models they were applied to. Table 4 presents
1∑
𝑛
the detailed hyperparameter configurations. The models investigated in
MAE = |𝑦 − 𝑦̂𝑖 |
this study include N-BEATS, 1D-CNN (abbreviated as CNN), LSTM, and 𝑛 𝑖=1 𝑖
CNN-LSTM. The 1D-CNN or CNN model is a variant of convolutional (9)
1∑
𝑛
neural networks specifically designed to operate on one-dimensional MSE = (𝑦 − 𝑦̂𝑖 )2
𝑛 𝑖=1 𝑖
data, in contrast to the standard 2D CNNs. The CNN-LSTM model
is a hybrid architecture that commences with a multi-layer 1D-CNN, Furthermore, this study examines the models’ capability to detect the
and its output is subsequently fed into a multi-layer LSTM model. An HABs alert levels in several future horizons, which can be regarded as
additional hyperparameter, Skipping Connections, is considered in the a multi-horizon classification problem. WHO considers different HABs
CNN-LSTM model, which determines whether the output of the initial vigilance and alert levels based on the intended use of water. Likewise,
1D-CNN network is concatenated with the input of the subsequent 1D- the implementation of these levels may vary according to national and
CNN network. The other hyperparameters listed in Table 4 are specific local regulations. To simplify the various alert scenarios and facilitate
to each architecture, while the hyperparameters under the ‘‘Common’’ comparisons between the different models, this study opted for a Chl-a
row are shared across all architectures. To ensure comprehensive cover- alert level of 10 μg∕L. This particular threshold falls within the current
age of the parameter space, the sets of hyperparameters were randomly range of the vigilance level for recreational waters and alert level 1
sampled from a uniform distribution. for drinking water [28]. Additionally, it aligns with the guidance level
The hyperparameters explored in this random search phase, along 1 outlined in the previous WHO 1999 edition [61]. It is important to
with their ranges, can be observed in Table 4. The architectures ex- highlight that the tool developed for this study can be fine-tuned based
plored include N-BEATS, 1D-CNN (abbreviated as CNN), LSTM, and on specific thresholds as needed.
CNN-LSTM (see Section 2 for further details about the architectures). Evaluation of this classification problem commonly employs the F1-
The CNN-LSTM model is a hybrid model that starts with a multi-layer score metric, which represents the harmonic mean of precision and
1D-CNN, and its output serves as the input for a multi-layer LSTM recall. The F1-score metric assesses whether the model’s predictions

9
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Table 4 For every model category, we explored up to 30 unique hyperparameter

Hyperparameters for each DL architecture used in the study, along with their
sets, resulting in a total of over 2000 trained models. An in-depth
respective data types and the range of values considered. The hybrid CNN-LSTM
architecture incorporates hyperparameters from both the CNN and LSTM architectures.
overview of the training methodology and model configurations can
Hyperparameters represented with square brackets indicate that they were sampled be referenced in the preceding Section 3.
from a uniform distribution for Integer types and a logarithmic uniform distribu- This section synthesizes the results from the evaluation of the mod-
tion for Real types. Values from ranges without specified intervals were uniformly els once trained and validated, and examines the practical implications
selected.
of the developed prediction system. The models were trained using
Architecture Parameter Type Range
data collected from the beach buoy over a period of two years, and
Dropout Real [0, 0.7] subsequently tested against data from the following year of both the
Common Batch normalization Integer [0, 1]
beach and dam buoys. This strategy was formulated to ascertain the
L2 regularization Real [0, 1]
generalizability of the models over different sampling points within
Number cells Integer [1, 200]
LSTM
Number layers Integer [0, 5]
the same water body. Notably, the N-BEATS model outperformed all
the other models evaluated in this study, showing superior accuracy
Number filters Integer 8, 16, 31, 64, 128
Kernel size Integer [2, 12]
in forecasting Chl-a levels and accurately identifying the critical WHO
CNN alert threshold for Chl-a concentrations.
Max pooling Integer [1, 2]
Number layers Integer [0, 8] The following subsections offer an in-depth analysis of model re-
CNN-LSTM Skipping connections Boolean True, False sults, categorized based on their predictive time intervals. Focusing
Number stacks Integer [1, 10] initially on the short-term models trained for 24-h predictions, a major-
Number of blocks per stack Integer [1, 10] ity of models consistently produced highly accurate predictions. In the
N-BEATS
Theta’s dimension Integer [1, 10] following section, we compare the initial 24-h predictions of the long-
Non-linear FC layers units Integer [32, 1024] range models, which provide forecasts for 168 h (or a week), with those
Non-linear FC layers per block Integer [1, 10]
Shared weights in stack Boolean True, False
of their daily counterparts. This comparison highlights a slight decrease
in the predictive accuracy of the long-term models. In the category
of long-range forecasts, the N-BEATS model stands out as particularly
noteworthy, achieving the lowest Mean Absolute Error (MAE) and
correctly identify the HAB alert level as compared to the original sen- approaching F1-scores of 70%, even on the third day of prediction.
sor. In this study, the F1-score metric was employed to select the best Notably, when these models were fine-tuned using beach buoy data and
models for both the Random Search with hyperparameter validation then applied to dam buoy data, they retained their effectiveness with
and the test dataset. This choice stems from the practical significance only a modest 10% decrease in their ability to identify the initial alert
of predicting future HAB alert in the environmental domain. However, threshold at this distinct location. In the final section, we reflect on the
metrics such as MAE and MSE were also utilized to enable experts to practical benefits and extensive versatility of the predictive approach
compare the performance of models on the test dataset. introduced in this study.
TP
Precision =
TP + FP 4.1. One-day ahead forecast
TP
Recall =
TP + FN (10) The selected DL models in this subsection have been trained with
2 ⋅ Precision ⋅ Recall the primary aim of predicting Chl-a values a day ahead at an hourly
𝐹 1-score =
Precision + Recall granularity. These models have been primarily assessed based on their
TP = True Positives, FP = False Positives, FN = False Negatives ability to predict the Chl-a value (evaluated using the MAE metric) as
well as their capability to identify the first level of HABs according
3.4. Experimental environment to WHO standards (evaluated using the F1-score metric) in the last
hour of the prediction (t + 24 h into the future). It should be noted
Deep learning models were trained and tested on an off-the-shelf that the models generated up to 24 values, each corresponding to an
computer with the following configurations: an Intel(R) Xeon(R) E5- hour from time t + 1 up to t + 24 in the future. Initially, the models
2630 v2 @ 2.60 GHz CPU, 64 GB of RAM, and a GeForce GTX 1080 Ti were filtered based on the validation dataset to identify first level alert
GPU. The software libraries used along with their respective versions 24 h into the future. For each model type, window size, and input
were as follows: Python 3.9.6, TensorFlow 2.4.1, Keras 2.4.0, Optuna features set, the optimal hyperparameters were selected. The results of
2.10.0, Pandas 1.3.1, and NumPy 1.18.5. these optimal hyperparameters, tested on both the beach and dam buoy
Please note that all the software utilized in this study is freely testing datasets, are presented in this section.
available to the public. The data used to conduct the experiments in this The testing phase results for the best models, which were selected
work can be accessed from the following repository: https://github.co based on the highest mean F1-score obtained on the validation dataset,
m/stanislavvakaruk/Chlorophyll_soft-sensor_machine_learning_models are shown in Fig. 7. The results presented in this figure indicate the
mean F1-score and mean MAE of the 24-h forecast (mean result of 24
4. Results and discussion values) for each buoy test dataset. As observed, three of four models
yielded high results, with a negligible difference in F1-score values
In this study, various DL models have been trained to predict (less than 1%). Specifically, the N-BEATS (83.5% mean F1-score), CNN
Chl-a levels over multiple time horizons and identify the preliminary (83.7% mean F1-score), and CNN-LSTM (84% mean F1-score) models
alarm threshold consistent with WHO guidelines for HAB control in achieved the best results for the beach buoy. Therefore, these models
freshwater bodies [28]. Specifically, these models were designed to were capable of forecasting and identifying Chl-a values at the alert
predict at both daily and weekly intervals, delivering forecasts for 24 h threshold in a more effective manner than the LSTM model (81% mean
in the daily scenario and 168 h for the weekly scenario (hourly data F1-score). The decrease of approximately 5% in mean F1-score from
granularity). Throughout the training, various past window lengths beach buoy to dam buoy suggests a strong generalization capability
were employed for each model type, namely N-BEATS, LSTM, CNN, of all the models. These three models were close in their predictions
and CNN+LSTM. These lengths spanned 12, 24, 72, 168, and 336 h. even in the intermediate values, as it can be seen in Fig. 8. However,
The endogenous variable, i.e. Chl-a, was used both in isolation and N-BEATS obtained the lowest values, and therefore the best results
alongside exogenous variables of temperature, pH, and conductivity. according to the mean MAE metric (1.069 in beach buoy and 1.100 in

10
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Table 5
Summary of the DL architecture, input, and output variables configurations explored in this study. For each of these
combinations, up to 20 random hyperparameter configurations were evaluated. Ultimately, up to 2000 deep learning
models were trained.
Parameter Range
Deep learning architecture N-BEATS, LSTM, CNN, CNN-LSTM
Input variables set Chl-a, Chl-a+Exo (Chlorophyll-a + Exogenous variables)
Input window size 12, 24, 72, 168, 336
Forecast window size 24, 168

Fig. 7. Mean F1 and MAE results for the best 1-day forecast model configurations. Both buoys used as sources of data are illustrated on the 𝑥-axis, whereas the mean F1-score
and mean MAE over 24 time steps are shown in the 𝑦-axis on the left and right figure, respectively. Each of the bars represents one of the models selected for comparison. In
addition, the exact value represented on the 𝑦-axis, the window and the variables used for the model are indicated on top of each bar. For the F1-score metric, higher values are
better. On the contrary, lower MAE values signify higher accuracy. The top-performing models for the mean F1-score and mean MAE over 24 time steps, averaged between the
two buoys, are CNN and N-BEATS, respectively.

Fig. 8. F1 score for the best 1-day forecast model configurations on each time step. Steps into the future and F1 are shown in 𝑦-axis and 𝑥-axis accordingly. The best configuration
per model is selected and assigned a different color line. Higher value is directly correlated to better accuracy. Beach buoy results are on the left and dam buoy on the right. The
best model on the last hour is CNN.

dam buoy), which implies its better capability to infer the exact value endogenous-only versions. In any case, these exogenous variables did
of Chl-a. The mean MAE of N-BEATS was 24% better than that obtained not seem to significantly differentiate these models from the others.
by the second best CNN model (1.333 MAE). This result became clear These results highlighted the superior performance of N-BEATS over
in Fig. 9, where N-BEATS’ line is clearly below the other lines even in conventional DL models in predicting day-ahead Chl-a values. N-BEATS
the intermediate values. not only achieved the minimal mean MAE but also showcased an
In terms of window size, all the best models used a window size F1-score with a difference of less than 1% from the top-tier model,
of 72 h, which points to better performance at this value. This results CNN-LSTM. Thus, this study suggests that the N-BEATS model could be
could be explained as the need for 3 days of past data to correctly the ideal model for the prediction of short-term Chl-a and identification
predict and identify abnormal Chl-a levels. With larger time windows, of alert levels based on WHO guidance.
the results start to deteriorate, which implies that the models need
to be tuned taking into account the amount of historical information.
4.2. One-week ahead forecast
Interestingly, the N-BEATS and CNN models obtained their best results
using only the endogenous variable (i.e. Chl-a) as input variable , unlike
the CNN-LSTM and LSTM models, which performed better considering One-week ahead forecasting is focused on predicting 168 timesteps
Chl-a and exogenous variables. The LSTM architecture, tailored for into the future. As previously mentioned, each of these time steps
sequential data, excels in discerning and assimilating the trend and pe- corresponds to one hour. Hence, the difficulty of this task is greater
riodicity inherent in the Chl-a level time series. As a result, the nuances than for one day forecasting and higher error is expected the further
of trends and periodicity become increasingly salient, overshadowing into the future it is predicted. Longer-term forecasting allows us to
the contributions of exogenous variables in extended forecasting sce- have a better idea of how Chl-a levels will evolve, evaluate possible
narios. Conversely, in short-term forecasting, the input from exogenous next steps and take earlier action. The experiments conducted for this
variables gains prominence, providing crucial information for achieving subsection were analogous to the day-ahead prediction experiments.
high accuracy in LSTM-based models predictions with respect to their The best model of each DL architecture was selected based on the

11
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 9. MAE for the best 1-day forecast model configurations on each time step. The steps into the future are represented in the 𝑥-axis, and MAE is shown in the 𝑦-axis. Each of
the best configurations per model are given a color. Lower value suggests higher accuracy. Graphs correspond to beach and dam from left to right. N-BEATS obtains the lowest
MAE on the 24th step.

Fig. 10. Comparison of the 1st day forecast for one-day and one-week ahead forecasting. The 𝑥-axis indicates the model to which each of the results refers and the mean F1 over
the first 24 time steps is shown at the top of each bar. The purple bars illustrate one-day ahead forecasting mean, while turquoise represents one-week ahead forecasting results.
10(a) shows the results for beach buoy while 10(b) shows the results for dam buoys. A higher F1 means a higher model exactness. CNN demonstrates superior performance for
both one-day and one-week ahead forecasting across both beach and dam buoy. However, N-BEATS exhibits the best performance for one-week ahead forecasting in both scenarios.

F1-score results obtained from the validation dataset and were subse- data structure in multi-horizon forecasting problems in a more effective
quently evaluated using the testing datasets from both the beach and way. In general, the use of long-term multi-horizon forecasting models
dam buoys (last year). Then, the best model configuration was chosen instead of short-term forecasting models results in a 3% decrease, which
for each DL model architecture, window size and input features set. can be considered a negligible cost for the flexibility of training only
Additionally, the one-week ahead forecasting models were compared one long-term forecasting model instead of potentially hundreds or
with the one-day ahead prediction models discussed in the previous thousands of short-term forecasting models.
subsection. To evaluate the forecasting reliability over a one-week ahead fore-
The first analysis in this section shows the flexibility of employing cast horizon, we analyzed the performance metrics for each day’s
a long-term multi-horizon forecasting model instead of a short-term predictions. As depicted in Fig. 11, the mean F1-score for the best
forecasting model. In a multi-horizon model, we can select the forecast configurations across different models is presented as the mean value
horizon, allowing us to seamlessly use a forecasting model designed of each day. Notably, N-BEATS stands our for its consistently superior
for one-week ahead predictions for one-day ahead forecasting. In this results, with an F1-score approaching 70% by the close of day 3. This
scenario, we only select the 24th prediction step of the long-term dominance is further emphasized in Fig. 12, where N-BEATS predom-
forecasting model from its 168 prediction steps. In order to compare inantly surpasses the observed performance of the other models. This
the long-term to the short-term forecasting model, we assess the 24th linear plot, however, highlights intermittent decreases in the model’s
step prediction of the long-term model alongside the last prediction performance. In particular, the Chl-a level showed less consistency
step (24th) of the short-term forecasting model, showcased in the during the day, resulting in less accurate predictions. There are fluc-
previous Section 4.1. In Fig. 10, the results for the best hyperparam- tuations in the level of Chl-a between the day and night time due to
eter configuration of each architecture on the 24th step in one-week the vertical migration of algae along the water column in search of
forecasting are shown in turquoise color, while the results for one-day sunlight. Even so, the N-BEATS model continued to show its robustness
ahead forecasting are shown in purple color. In general terms, N-BEATS by exhibiting lower MAE values compared to the other models (Fig. 13).
obtains the best results for the first day in one-week ahead forecasting This advantage is further evidenced in Fig. 14, where the trajectory
on both buoys with a F1-score of 80.9% on beach buoys and 76.3% on of N-BEATS (indicated by a blue line) persistently ranks below the
dam buoy, respectively. Although the short-term version of N-BEATS competing models throughout the observed duration.
differs less than 1% from the CNN and CNN+LSTM models, in the As depicted in the figure presented above, N-BEATS consistently
long-term version it is more than 3% higher than them in both beach outperforms the other models. By the end of day 2, particularly on
and dam buoys. The sequence based architectures such as N-BEATS the dam buoy, N-BEATS approaches an F1-score of approximately
and LSTM outperform the other models as they exploit the temporal 70%. As elaborated in Section 4.1, a typical decline in generalization

12
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 11. Daily mean of the F1-score for the best one-week forecasting model configurations. Days in the forecasted week are shows in the 𝑥-axis, while the mean F1 per day is
shown in 𝑦-axis. Each color bar is assigned to the best configuration for each model. Beach buoy is represented in (a) and dam buoy is represented in (b). Higher values indicate
better alarm 1 classification. N-BEATS obtains the best overall results.

Fig. 12. F1 for the best one-week ahead forecasting model configurations. Time steps are indicated in the 𝑥-axis and F1 is shown in 𝑦-axis. The best model configurations are
assigned a unique color line. Higher F1 score equates to better model exactness. Beach buoy results are shown in (a) and dam buoy results on the (b). The best model in the
majority of horizons including the last 50 h is N-BEATS.

capability, roughly around 10%, is observed on the second buoy, even the last observed value as the forecasting for all future timesteps. The
over extended periods. However, the results in Fig. 12 reinforces the table shows optimal configurations for each buoy, prediction horizon,
supremacy of N-BEATS, consistently placing it above its counterparts and associated DL architecture, contingent on the feature set and the
for the dam buoy. A similar trend is observed when evaluating the model’s input window size. It should be noted that, the configurations
MAE as shown in Figs. 13 and 14. Notably, Fig. 14 reveals a mere for the dam buoy align with the beach buoy, since predictions for the
5% difference in MAE between the beach and dam buoy for N-BEATS, dam buoy leveraged the best-performing models initially trained on
signifying an even more formidable generalization capability than what beach buoy data. Each optimal configuration features presents average
is inferred from the F1-score. Broadening the lens to a long-term view, MAE, MSE, and F1-score values, computed both daily and weekly (for
N-BEATS’ prowess in generalization becomes undeniably apparent. more details about the metrics see the Section 3.3). While models
With reductions of approximately 10% and 5% in F1-score and MAE, with daily forecast horizons cannot directly produce weekly forecasts,
respectively, N-BEATS demonstrates its ability to estimate the level of those designed for weekly predictions can offer insights for their first
Chl-a at various locations in the water body.
predicted day. The table also shows inference speeds, with all models
registering under one second per prediction. As previously emphasized,
4.3. Significance of deep learning models for chlorophyll-a forecasting
N-BEATS consistently outperforms other models in terms of MAE and
MSE, indicating its superiority in closely approximating future Chl-a
The one-day and one-week ahead forecasting results demonstrate
values. In numerous instances, N-BEATS demonstrates a substantially
that our proposed methodology produced DL model configurations
reduced MSE, suggesting fewer pronounced errors relative to other
capable of predicting Chl-a values over time, maintaining high accuracy
in identifying the alert threshold at 10 μg∕L up to the third day. models. In relation to the identification of the 10 μg∕L Chl-a, N-BEATS
Accurate predictions threshold based on WHO guidance are crucial stands out, particularly for long-term, one-week ahead forecasts, and
for enabling informed decision-making related to sampling, protection, only marginally diverges from top models for daily forecasts by less
and risk management in water bodies affected by algal blooms. In this than 1%. These long-term forecasting models hold pronounced value
subsection, we analyze the results by focusing on the time windows due to their inherent flexibility of forecast selection. N-BEATS, when
and prediction speeds. False negatives and positives are then examined trained on a one-week ahead forecast horizon, is notably competitive
and modifications to the prediction threshold are evaluated to optimize at one-day ahead forecasts. As depicted in Fig. 12, N-BEATS obtained
detection of WHO-defined alert. a F1-score approaching 70% on the third day of prediction for the
Table 6 compiles the findings presented in the prior sections, pro- beach buoy, i.e., the one used to train the model. Additionally, its
viding a comparative analysis with the results obtained using the robust generalization capabilities shine through, as evidenced by nearly
Random Walk method as a benchmark. This naïve approach utilizes matching performance on the second day for the dam buoy situated

13
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 13. Daily average of the MAE for the best one-week ahead forecasting model configurations. Days are shown on the 𝑥-axis. The 𝑦-axis illustrates the mean MAE of the
corresponding day. A unique color bar is assigned to the best configuration for each distinct architecture. (a) shows the results for beach buoy and (b) the results for dam buoy.
Lower values suggest better forecasting accuracy. N-BEATS obtains the best overall results.

Fig. 14. MAE on each forecasted time step for the best one-week ahead forecasting model configurations. Time steps are indicated in the 𝑥-axis and F1 is shown in 𝑦-axis. Unique
colors are assigned to the best model configurations. Lower MAE equates to better forecasting accuracy. (a) shows the results for the beach buoy while (b) shows the results of
dam buoy. The best model on the on all predicted hours is N-BEATS.

Table 6
Summary of results categorized by buoy testing dataset, forecast horizon, and architecture. The table represents optimal trained models based on the best feature set and input
window size. Configurations for the dam buoy are consistent with those of the beach buoy. Displayed metrics include average MAE, MSE, and F1 values for both daily and weekly
forecast horizons. It is worth noting that while models with a daily horizon are not designed for weekly predictions, those with a weekly horizon offer results for the first day.
Additionally, inference time in seconds are provided.
Buoy Horizon Architecture Features Window MAE (day) MSE (day) F1 (day) MAE (week) MSE (week) F1 (week) Inference time (s)
N-BEATS Chl-a w72 1.069 4.394 0.835 – – – 0.0025
LSTM Chl-a+Exo w72 2.024 6.866 0.810 – – – 0.0052
Day CNN Chl-a w72 1.333 4.439 0.837 – – – 0.0026
CNN-LSTM Chl-a+Exo w72 1.355 4.665 0.840 – – – 0.008
Random Walk Chl-a w12 1.319 8.530 0.774 – – – –
Beach
N-BEATS Chl-a w12 1.124 4.410 0.809 2.098 13.354 0.638 0.0023
LSTM Chl-a w72 1.871 6.650 0.792 2.806 16.805 0.599 0.0074
Week CNN Chl-a w24 2.030 6.758 0.774 3.247 18.213 0.607 0.0045
CNN-LSTM Chl-a w72 1.527 6.191 0.777 2.342 15.745 0.619 0.014
Random Walk Chl-a w12 1.257 7.688 0.769 2.027 17.568 0.636 –
N-BEATS Chl-a w72 1.100 4.303 0.767 – – – 0.0026
LSTM Chl-a+Exo w72 2.317 8.293 0.748 – – – 0.0052
Day CNN Chl-a w72 1.350 4.405 0.782 – – – 0.0025
CNN-LSTM Chl-a+Exo w72 1.478 4.877 0.776 – – – 0.0077
Random Walk Chl-a w12 1.363 8.326 0.699 – – – –
Dam
N-BEATS Chl-a w12 1.165 4.697 0.763 2.158 12.279 0.549 0.0023
LSTM Chl-a w72 1.878 7.705 0.744 2.762 14.885 0.523 0.0075
Week CNN Chl-a w24 2.075 7.369 0.730 3.245 16.832 0.524 0.0044
CNN-LSTM Chl-a w72 1.627 7.781 0.718 2.429 14.400 0.528 0.0139
Random Walk Chl-a w12 1.322 7.994 0.702 2.135 17.821 0.531 –

on the opposing lake end. The conclusions were validated applying for each dataset and forecasting horizon. As previously concluded, N-
a Model Confidence Set (MCS) [62] to the non-conditional forecasts BEATS consistently demonstrates superior performance within every
obtained from the endogenous models. This procedure narrowed down model confidence set. Detailed results of this validation can be found
the initial set of models to a subset with equivalent predictive ability in Appendix.

14
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Fig. 15. Assessment of alert detection based on Chl-a threshold variations. This figure juxtaposes detection thresholds from N-BEATS model predictions with the selected Chl-a
threshold (10 μg∕L). The 𝑥-axis enumerates the alternative thresholds in μg∕L. On the 𝑦-axis, the false positive rates, represented by the blue line and based on 7000 possible
values from the testing dataset, and the false negative rates, indicated by the orange line and based on 1000 possible samples, are visualized in the context of 10 μg∕L Chl-a alert
detection. Fig. 15(a) depicts the performance of the model trained with a one-day forecast horizon for the beach buoy. Similarly, Fig. 15(b) portrays results for the dam buoy
under the same forecast horizon. For models trained with a one-week ahead forecast horizon, Fig. 15(c) provides insights for the beach buoy, and Fig. 15(d) highlights findings
for the dam buoy.

In our study, as shown in Table 6, a distinct preference emerged On the other hand, predicting algal blooms presents an inherent
for adopting a 72-h input time window when making one-day ahead challenge: balancing false negatives (i.e. HABs that do not trigger the
forecasts. However, for week-long forecasts, no such clear trend is alert) against false positives (i.e. erroneously anticipated HABs). The
observed. For projections spanning a week, DL models appeared to be first pose public health risks, while the latter could have economic
more effective at discerning trend and periodicity patterns intrinsic to consequences, especially if they cause unnecessary restrictions of lakes
the time series. These patterns, more prominent over a week than a used for tourism, drinking water supply or for industrial activities that
single day, increased the model’s performance metrics during training depend on clean water (e.g., hydro-power production) [28]. Fig. 15
on a week-long forecast horizon. Models such as N-BEATS excel at iden- contrasts these errors, with blue denoting false positive percentages
tifying trend and periodicity patterns without necessitating exhaustive (based on 7000 possible instances of the testing dataset) and orange
input window size. On the other hand, for daily forecasts, the relevance representing false negatives (based on 1000 possible instances). It was
of trend or periodicity patterns of the time series was reduced, leading evident that daily forecasts manifested fewer discrepancies (both false
to a direct prediction challenge. Here, models took advantage of the positive and negatives) than their weekly counterparts. This is likely
larger data sets, often resorting to the largest possible window size because long-term forecasts inherently face greater complexities.
and occasionally integrating exogenous variables. Interestingly, the fact Typically, there is a tendency that, as the alert threshold increases,
that the models used exclusively the previous Chl-a values to predict false negatives increase and false positives decrease, and vice versa.
Generally, at the 10 μg∕L threshold, false positives were surpassed
suggests that, at least in certain water bodies such as the As Conchas
by false negatives (Fig. 15). In other words, as previously explained,
reservoir, the use of Chl-a sensors for model training and alerting
the potential economic impact (related to the false positives) was
would be sufficient. This simplifies and reduces the cost of AHFM
outweighed by health risk (related to the false negatives). However,
system design and operation, since the fewer sensors used, the less
the model’s adaptability empowers users to recalibrate this balance
investment and maintenance of the nodes. Furthermore, the rapid
by opting for an alternate threshold in line with their specific needs.
prediction capabilities presented by these models pave the way for
For instance, in water bodies for drinking water purposes, considering
further reductions in associated hardware expenses.
lower thresholds could refine detection fidelity. Pairing such threshold
All the models, notably those based on the N-BEATS architecture, adjustments with the optimal prediction horizon can fine-tune both the
are capable of producing predictions in as little as every 2 ms, a speed prediction quality and the balance of false indicators. Yet, irrespective
that vastly outpaces the conventional 15-min data collection interval. of threshold nuances, it is of paramount importance to corroborate
Leveraging the robust generalization ability of these models across any generated alerts with rigorous laboratory assessments for harmful
different buoys, N-BEATS can achieve predictions for Chl-a levels and species and toxin concentration, safeguarding both precision and public
the WHO alert thresholds for an impressive array of up to 300,000 sam- health.
pling points within each 15-min span, given the hardware discussed in
this study (refer to Section 3.4). In the future context of environmental 5. Conclusions and future work
digital twins, where virtual replicas of physical systems or processes
are created, the fast predictive capabilities of N-BEATS are perfectly The study delves into predicting Harmful Algal Blooms (HABs),
aligned. This suggests that there is a substantial realm of opportunities focusing on Chlorophyll-a levels and their threats to environmental,
to reduce computational costs for predictive models, potentially leading animal, and human health in the Anthropocene era. With the ap-
to an embedded solution within the buoy alongside the sensors. plication of novel Deep Learning (DL) architectures such as Neural

15
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Basis Expansion Analysis for Interpretable Time Series Forecasting (N- constrain the utilization of predicted Chlorophyll-a levels for other
BEATS), it improves Chlorophyll-a level forecasting with high temporal analytical purposes, such as examining the evolution and dynamics of
resolution and extends the horizon to one week. Utilizing a publicly HABs.
available dataset from the As Conchas reservoir, the study trains and
evaluates an extensive array of models, totaling up to 2400 distinct Code availability
configurations, with N-BEATS standing out for its ability to capture
temporal dependencies efficiently. This DL architecture demonstrates The code and data used to conduct the experiments presented in
superior long-term forecasting while also proving effective for short- this work can be found in the following publicly accessible repos-
term predictions, suggesting its potential as a substitute for short-term itory: https://github.com/stanislavvakaruk/Chlorophyll_soft-sensor_ma
models with minimal loss in precision. chine_learning_models.
The main conclusions of this study are the following:
CRediT authorship contribution statement
• Innovative Deep Learning Application: By employing advanced
DL architectures, specifically N-BEATS, for predicting Chlor- Silvia Martín-Suazo: Writing – review & editing, Writing – orig-
ophyll-a levels, a high temporal resolution and an extension of inal draft, Visualization, Validation, Software, Resources, Methodol-
the forecasting horizon to one-week are achieved. This approach ogy, Investigation, Formal analysis, Data curation, Conceptualization.
shows promise in replacing short-term forecasting models with Jesús Morón-López: Writing – review & editing, Writing – origi-
minimal loss of accuracy. nal draft, Validation, Resources, Methodology, Investigation, Funding
• Practical Advancements in HAB Management: A novel method- acquisition, Formal analysis, Conceptualization,. Stanislav Vakaruk:
ology for predicting initial alarm levels of HABs within a multi- Writing – review & editing, Writing – original draft, Visualization,
horizon Chlorophyll-a level forecasting framework is proposed. Validation, Supervision, Software, Project administration, Methodol-
This methodology is aligned with World Health Organization ogy, Investigation, Formal analysis, Data curation, Conceptualization.
(WHO) guidelines and offers practical advantages for refined HAB Amit Karamchandani: Writing – review & editing, Writing – origi-
management and mitigation initiatives. nal draft, Validation, Investigation, Formal analysis, Conceptualization.
• Effectiveness of Supplementary Variables: The integration of sup- Juan Antonio Pascual Aguilar: Writing – review & editing, Writing –
plementary variables such as water temperature, pH, and con- original draft, Validation, Resources, Formal analysis, Data curation,
ductivity has been found to not substantially improve predictive Conceptualization. Alberto Mozo: Writing – review & editing, Writing
accuracy, affirming the validity of simpler models and challeng- – original draft, Validation, Supervision, Software, Project adminis-
ing assumptions about increased complexity leading to enhanced tration, Investigation, Funding acquisition, Conceptualization,. Sandra
performance. Gómez-Canaval: Writing – review & editing, Writing – original draft,
• Model Portability: Models trained on data from one buoy exhib- Visualization, Validation, Methodology, Investigation, Formal analysis,
ited minimal degradation in precision when applied to another Conceptualization. Meritxell Vinyals: Writing – review & editing, Val-
buoy within the same water body, indicating potential for stream- idation, Methodology, Investigation, Formal analysis, Conceptualiza-
lined model training methodologies and broadened utility across tion. Juan Manuel Ortiz: Writing – review & editing, Writing – original
various water body locations. draft, Validation, Resources, Funding acquisition, Conceptualization.
• Refined Alarm Level Inference: Interesting opportunities have
been studied to refine predictive model alarm level inference by Declaration of competing interest
adjusting prediction thresholds, ensuring more timely and precise
alarm identification, and ultimately preserving human health. The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
As future work, building upon the insights obtained from this study, influence the work reported in this paper.
there are multiple directions for future research to enhance the fore-
casting and understanding of HABs. A primary area of focus should Data availability
be to increase data collection efforts, including a variety of aquatic
environments and capturing a wider spectrum of variables that have The dataset used in this study is publicly available and can be ac-
known associations with HAB occurrences. This augmented dataset cessed at https://github.com/stanislavvakaruk/Chlorophyll_soft-sensor_
could potentially enable more complex models to accurately determine machine_learning_models. The dataset includes measurements of Chl-
higher-risk alarm levels, which denotes a more severe threat. The a, pH, water temperature, electrical conductivity, and other relevant
extension of the forecast horizon to span one month may provide parameters collected from the As Conchas reservoir in the Miño-
valuable insights into long-term HAB trends, although this would neces- Sil River Basin District of Galicia, NW Spain. Detailed information
sitate access to extensive water quality and climate data, especially in regarding buoy location and data acquisition can be found in Section
relation to nutrient inputs such as nitrogen and phosphorus. Employing 3.1.
broader temporal granularities, such as several hours or a full day,
would also afford a more comprehensive view of short-term HAB Acknowledgments
patterns, contributing to the development of response strategies and
ultimately improving our ability to manage and mitigate the impacts of This work has been supported in part by Fundación Biodiversidad,
HABs. The implementation of transfer learning methodologies presents Spain, the Spanish Ministry for Ecological Transition and the Demo-
a promising direction for future work. This approach, which uses pre- graphic Challenge with grant CA-CC-2018 (CianoMOD), and partially
trained models on similar tasks or datasets, could simplify the training by the Spanish Ministry of Economy, Industry and Competitiveness
process, facilitating a quicker and more efficient path to accurate HAB (MINECO), Spain, co-financed by the European Union (FEDER), under
forecasting models by capitalizing on previously learned patterns and the grant RTC-2016-5087-2 (CIANOALERT). The co-authors thank the
knowledge. Additionally, there is potential for developing specialized Miño-Sil Hydrographic Confederation for their support in the develop-
classification models explicitly designed for HAB alarm detection. Such ment of the project and for facilitating the access to the water quality
models could bypass the forecasting stage altogether, enabling direct data. This work was partially supported by Universidad Politécnica de
alarm-level classification. While this might yield faster and potentially Madrid, Spain under the project MLSEC with reference RP2161220029.
more precise assessments, it would also limit the flexibility of the ap- The authors would like to thank M.Sc. student Miriam Guindel Gómez
proach compared to the methodology outlined in this study, as it would for her help with the design and coding of the preliminary experiments.

16
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

Table A.1
Results of the MCS relative to the mean MAE of all timesteps on the unconditional forecasts. Each forecasting horizon’s models were evaluated
using data from both buoys. The column denoted as 𝑆 represents the critical value threshold, while the critical value indicates the predictive
equivalence of the actual confidence set. The p-value reflects the statistical significance of these results. Additionally, the excluded model and
resulting confidence set for each step are detailed.
Dataset Horizon 𝑆 Critical value p-value Removed Model confidence set
182.910 9.210 0.000 CNN [N-BEATS, Random Walk]
Day
16.770 6.635 0.000 Random Walk [N-BEATS]

Beach 336.190 13.277 0.000 CNN [N-BEATS, LSTM, CNN-LSTM, Random Walk]
142.448 11.345 0.000 LSTM [N-BEATS, CNN-LSTM, Random Walk]
Week
30.469 9.210 0.000 CNN-LSTM [N-BEATS, Random Walk]
1.211 6.635 0.271 – [N-BEATS, Random Walk]
93.408 9.210 0.000 Random Walk [N-BEATS, CNN]
Day
52.270 6.635 0.000 CNN [N-BEATS]

Dam 284.010 13.277 0.000 CNN [N-BEATS, LSTM, CNN-LSTM, Random Walk]
137.200 11.345 0.000 LSTM [N-BEATS, CNN-LSTM, Random Walk]
Week
31.968 9.210 0.000 CNN-LSTM [N-BEATS, Random Walk]
0.106 6.635 0.745 – [N-BEATS, Random Walk]

Table A.2
Results of the MCS relative to the mean MSE of all timesteps on the unconditional forecasts. Each forecasting horizon’s models were evaluated using
data from both buoys. The column denoted as 𝑆 represents the critical value threshold, while the critical value indicates the predictive equivalence of
the actual confidence set. The p-value reflects the statistical significance of these results. Additionally, the excluded model and resulting confidence set
for each step are detailed.
Dataset Horizon 𝑆 Critical value p-value Removed Model confidence set
9.510 9.210 0.009 Random Walk [N-BEATS, CNN]
Day
0.048 6.635 0.827 – [N-BEATS, CNN]

Beach 86.763 13.277 0.000 CNN [N-BEATS, LSTM, CNN-LSTM, Random Walk]
15.714 11.345 0.001 Random Walk [N-BEATS, LSTM, CNN-LSTM]
Week
13.932 9.210 0.001 LSTM [N-BEATS, CNN-LSTM]
2.552 6.635 0.110 – [N-BEATS, CNN-LSTM]
Day 3.337 9.210 0.189 – [N-BEATS, CNN, Random Walk]
73.515 13.277 0.000 Random Walk [N-BEATS, LSTM, CNN, CNN-LSTM]
Dam
108.245 11.345 0.000 CNN [N-BEATS, LSTM, CNN-LSTM]
Week
17.186 9.210 0.000 LSTM [N-BEATS, CNN-LSTM]
8.763 6.635 0.003 CNN-LSTM [N-BEATS]

Appendix References

[1] Maciej Zalewski, Edyta Kiedrzyńska, Iwona Wagner, Katarzyna Izydorczyk,

Joanna Mankiewicz Boczek, Tomasz Jurczak, Kinga Krauze, Piotr Frankiewicz,
A.1. Model confidence set results Małgorzata Godlewska, Adrianna Wojtal-Frankiewicz, Małgorzata Łapińska, Mag-
dalena Urbaniak, Agnieszka Bednarek, Zbigniew Kaczkowski, Ilona Gągała,
Liliana Serwecińska, Sebastian Szklarek, Renata Włodarczyk-Marciniak, Arnoldo
The results presented in Section 4 were subsequently validated Font-Nájera, Elżbieta Mierzejewska, Małgorzata Połatyńska-Rudnicka, Kamila
through a Model Confidence Set (MCS). This procedure aims to obtain a Belka, Paweł Jarosiewicz, Ecohydrology and adaptation to global change, Eco-
hydrol. Hydrobiol. 21 (3) (2021) 393–410, http://dx.doi.org/10.1016/j.ecohyd.
set of models with equal predictive ability by conducting several rounds
2021.08.001.
of exclusions. Tables A.1 and A.2 provide an overview of each exclusion [2] Jeff C. Ho, Anna M. Michalak, Nima Pahlevan, Widespread global in-
step until the formation of the final confidence set. This procedure crease in intense lake phytoplankton blooms since the 1980s, Nature 574
(7780) (2019) 667–670, http://dx.doi.org/10.1038/s41586-019-1648-7, ISBN:
was performed for each forecasting horizon and buoy. Furthermore,
4158601916 Publisher: Springer US.
the analysis included both MAE and MSE metrics. Only models uti- [3] Zorica Svirčev, Dijana Lalić, Gorenka Bojadžija Savić, Nada Tokodi, Damjana
lizing Chl-a as an input feature were considered, as conditional and Drobac Backović, Liang Chen, Jussi Meriluoto, Geoffrey A. Codd, Global geo-
non-conditional models cannot be compared. On the first step of the graphical and historical overview of cyanotoxin distribution and cyanobacterial
poisonings, in: Archives of Toxicology, Vol. 93, Springer Berlin Heidelberg,
procedure, the critical value and 𝑆 of the initial set are calculated. If the 2019, pp. 2429–2481, http://dx.doi.org/10.1007/s00204-019-02524-4, ISSN:
critical value exceeds the threshold 𝑆, it signifies a substantial differ- 0340-5761 Number: 9.
ence between the models, thereby necessitating the exclusion step. The [4] Anna M. Michalak, Study role of climate change in extreme threats to water
quality, 2016, http://dx.doi.org/10.1038/535349a, URL http://go.nature.com/
𝑝-value quantifies the probability of obtaining the difference measured
29ye4ul. ISSN: 14764687 Number: 7612 Pages: 349–350 Publication title: Nature
by 𝑆 or an even larger difference, under the assumption of no true Volume: 535.
distinction between the models (i.e., when the null hypothesis holds). A [5] Cihelio Alves Amorim, Ariadne do Nascimento Moura, Ecological impacts of
freshwater algal blooms on water quality, plankton biodiversity, structure, and
𝑝-value below a threshold, set at 𝛼 = 0.01, signifies the rejection of the
ecosystem functioning, Sci. Total Environ. 758 (2021) 143605, http://dx.doi.org/
null hypothesis, which indicates no substantial difference between the 10.1016/j.scitotenv.2020.143605, Publisher: Elsevier B.V..
models under comparison. If a substantial difference is observed, pair- [6] Mulalo Mutoti, Jabulani Gumbo, Afam Israel Obiefuna Jideani, Occurrence
wise comparisons of models using the multivariate Diebold–Mariano of cyanobacteria in water used for food production: A review, Phys. Chem.
Earth A/B/C 125 (2022) 103101, http://dx.doi.org/10.1016/J.PCE.2021.103101,
test are conducted and the model with highest anticipated future loss Publisher: Pergamon.
is excluded. This exclusion process continues until a subset of models [7] Elisa Berdalet, Lora E. Fleming, Richard Gowen, Keith Davidson, Philipp Hess,
with equal predictive capabilities is achieved, or until only one model Lorraine C. Backer, Stephanie K. Moore, Porter Hoagland, Henrik Enevoldsen,
Marine harmful algal blooms, human health and wellbeing: Challenges and
remains in the subset. opportunities in the 21st century, J. Mar. Biol. Assoc. UK 96 (1) (2016) 61–91,
http://dx.doi.org/10.1017/S0025315415001733.

17
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

[8] Isabella Sanseverino, Diana Conduto, Algal bloom and its economic impact, ISBN: [26] Jiabao Wen, Jiachen Yang, Yang Li, Liqing Gao, Harmful algal bloom warning
978-92-79-58101-4, 2016, http://dx.doi.org/10.2788/660478. based on machine learning in maritime site monitoring, Knowl.-Based Syst. 245
[9] The water crisis does not stop in New York, Nat. Water 1 (4) (2023) 301, (2022) 108569, http://dx.doi.org/10.1016/J.KNOSYS.2022.108569, Publisher:
http://dx.doi.org/10.1038/s44221-023-00078-5. Elsevier.
[10] Catherine E. Richards, Asaf Tzachor, Shahar Avin, Richard Fenner, Rewards, [27] Jin Hwi Kim, Jae Ki Shin, Hankyu Lee, Dong Hoon Lee, Joo Hyon Kang,
risks and responsible deployment of artificial intelligence in water systems, Nat. Kyung Hwa Cho, Yong Gu Lee, Kangmin Chon, Sang Soo Baek, Yongeun Park,
water (2023) 1–11, http://dx.doi.org/10.1038/s44221-023-00069-6, URL https: Improving the performance of machine learning models for early warning of
//www.nature.com/articles/s44221-023-00069-6. Publisher: Nature Publishing harmful algal blooms using an adaptive synthetic sampling method, Water
Group. Res. 207 (2021) 117821, http://dx.doi.org/10.1016/J.WATRES.2021.117821,
[11] Silvia Beatriz Alves Rolim, Bijeesh Kozhikkodan Veettil, Antonio Pedro Vieiro, Publisher: Pergamon.
Anita Baldissera Kessler, Clóvis Gonzatti, Remote sensing for mapping algal [28] Ingrid Chorus, Martin Welker (Eds.), Toxic Cyanobacteria in Water. A Guide
blooms in freshwater lakes: a review, Environ. Sci. Pollut. Res. 30 (8) (2023) to Their Public Health Consequences, Monitoring and Management, second ed.,
19602–19616, http://dx.doi.org/10.1007/S11356-023-25230-2, URL https:// CRC Press, Geneva, 2021.
link.springer.com/article/10.1007/s11356-023-25230-2. Publisher: Springer. [29] María Castrillo, Álvaro López García, Estimation of high frequency nutrient con-
[12] Haobin Cen, Jiahan Jiang, Guoqing Han, Xiayan Lin, Yu Liu, Xiaoyan Jia, Qiyan
centrations from water quality surrogates using machine learning methods, Water
Ji, Bo Li, Applying deep learning in the prediction of chlorophyll-a in the east
Res. 172 (2020) 115490, http://dx.doi.org/10.1016/j.watres.2020.115490, URL
china sea, Remote Sens. 14 (21) (2022) http://dx.doi.org/10.3390/rs14215461.
https://www.sciencedirect.com/science/article/pii/S0043135420300269.
[13] Moein Izadi, Mohamed Sultan, Racha El Kadiri, Amin Ghannadi, Karem Abdel-
[30] Hiroshi Yajima, Jonathan Derot, Application of the Random Forest model for
mohsen, A remote sensing and machine learning-based approach to forecast the
chlorophyll- a forecasts in fresh and brackish water bodies in Japan, using
onset of harmful algal bloom, Remote Sens. 13 (19) (2021) http://dx.doi.org/
multivariate long-term databases, J. Hydroinform. 20 (2017) jh2017010, http:
10.3390/rs13193863.
//dx.doi.org/10.2166/hydro.2017.010.
[14] Hamed Karimian, Jinhuang Huang, Youliang Chen, Zhaoru Wang, Jinsong
Huang, A novel framework to predict chlorophyll-a concentrations in water [31] Bing Li, Guishan Yang, Rongrong Wan, Georg Hörmann, Jiacong Huang, Nicola
bodies through multi-source big data and machine learning algorithms, Environ. Fohrer, Lu Zhang, Combining multivariate statistical techniques and random
Sci. Pollut. Res. (0123456789) (2023) http://dx.doi.org/10.1007/s11356-023- forests model to assess and diagnose the trophic status of Poyang Lake in China,
27886-2, ISBN: 1135602327886 Publisher: Springer Berlin Heidelberg. Ecol. Indic. 83 (2017) 74–83, http://dx.doi.org/10.1016/j.ecolind.2017.07.033,
[15] Ting Zhou, Yan Li, Bo Jiang, Juha M. Alatalo, Chen Li, Cheng Ni, Tracking spatio- URL https://www.sciencedirect.com/science/article/pii/S1470160X17304466.
temporal dynamics of harmful algal blooms using long-term MODIS observations [32] Xue Li, Jian Sha, Zhong-Liang Wang, Application of feature selection and
of Chaohu Lake in China from 2000 to 2021, Ecol. Indic. 146 (2023) 109842, regression models for chlorophyll-a prediction in a shallow lake, Environ. Sci.
http://dx.doi.org/10.1016/J.ECOLIND.2022.109842, Publisher: Elsevier. Pollut. Res. 25 (2018) http://dx.doi.org/10.1007/s11356-018-2147-3.
[16] Andrea Vander Woude, Steve Ruberg, Thomas Johengen, Russ Miller, Dack [33] Jingjing Xia, Jin Zeng, Environmental factor assisted chlorophyll-a prediction
Stuart, Spatial and temporal scales of variability of cyanobacteria harmful algal and water quality eutrophication grade classification: a comparative analysis of
blooms from NOAA GLERL airborne hyperspectral imagery, J. Gt. Lakes Res. multiple hybrid models based on a SVM, Environ. Sci.: Water Res. Technol. 7
45 (3) (2019) 536–546, http://dx.doi.org/10.1016/j.jglr.2019.02.006, Publisher: (6) (2021) 1040–1049, http://dx.doi.org/10.1039/D0EW01110J, Publisher: The
International Association of Great Lakes Research. Royal Society of Chemistry.
[17] Katja Dörnhöfer, Philip Klinger, Thomas Heege, Natascha Oppelt, Multi-sensor [34] Li Wang, Xiaoyi Wang, Xuebo Jin, Jiping Xu, Huiyan Zhang, Jiabin Yu, Qian
satellite and in situ monitoring of phytoplankton development in a eutrophic- Sun, Chong Gao, Lingbin Wang, Analysis of algae growth mechanism and
mesotrophic lake, Sci. Total Environ. 612 (2018) 1200–1214, http://dx.doi.org/ water bloom prediction under the effect of multi-affecting factor, Saudi J. Biol.
10.1016/j.scitotenv.2017.08.219. Sci. 24 (3) (2017) 556–562, http://dx.doi.org/10.1016/j.sjbs.2017.01.026, URL
[18] Rafael Marce, Glen George, Paola Buscarinu, Melania Deidda, Julita Dunalska, https://www.sciencedirect.com/science/article/pii/S1319562X17300359.
Elvira de Eyto, Giovanna Flaim, Hans-Peter Grossart, Vera Istvanovics, Mirjana [35] Fang Lu, Zhi Chen, Wenquan Liu, Hongbo Shao, Modeling chlorophyll-a concen-
Lenhardt, Enrique Moreno-Ostos, Biel Obrador, Ilia Ostrovsky, Donald C Pierson, trations using an artificial neural network for precisely eco-restoring lake basin,
Jan Potuz, Sandra Poikane, Karsten Rinke, Sara Rodríguez-Mozaz, Peter A Staehr, Ecol. Eng. 95 (2016) 422–429, http://dx.doi.org/10.1016/j.ecoleng.2016.06.072,
Guido Waajen, Gesa A Weyhenmeyer, Kathleen C Weathers, Mark Zion, Bas W URL https://www.sciencedirect.com/science/article/pii/S0925857416303974.
Ibelings, Eleanor Jennings, Automatic high frequency monitoring for improved [36] Gooyong Lee, Faridah Othman, Shaliza Ibrahim, Min Jang, Determination of
lake and reservoir management, 50 (2016) 10780–10794, http://dx.doi.org/10. the forecasting-model parameters by statistical analysis for development of
1021/acs.est.6b01604. algae warning system, Desalin. Water Treat. 57 (55) (2016) 26773–26782,
[19] Jesus Moron-Lopez, Cristina Rodriguez-Sanchez, Francisco Carreno, Joaquin Va- http://dx.doi.org/10.1080/19443994.2016.1190106, Publisher: Taylor & Francis
quero, Angel G. Pompa-Pernia, Myriam Mateos-Fernandez, Juan Antonio Pascual tex.eprint: https://doi.org/10.1080/19443994.2016.1190106.
Aguilar, Implementation of smart buoys and satellitebased systems for the remote [37] P.J. García Nieto, E. García-Gonzalo, J.R. Alonso Fernández, C. Díaz Muñiz,
monitoring of harmful algae bloom in inland waters, IEEE Sens. J. (2020) 1–8, Water eutrophication assessment relied on various machine learning techniques:
http://dx.doi.org/10.1109/JSEN.2020.3040139. A case study in the Englishmen Lake (Northern Spain), Ecol. Model. 404 (2019)
[20] Abhiram S.P. Pamula, Hamed Gholizadeh, Mark J. Krzmarzick, William E. 91–102, http://dx.doi.org/10.1016/j.ecolmodel.2019.03.009, URL https://www.
Mausbach, David J. Lampert, A remote sensing tool for near real-time monitoring sciencedirect.com/science/article/pii/S0304380019301061.
of harmful algal blooms and turbidity in reservoirs, JAWRA J. Am. Water
[38] Sangmok Lee, Donghyun Lee, Improved prediction of harmful algal blooms in
Resour. Assoc. (2023) http://dx.doi.org/10.1111/1752-1688.13121, Publisher:
four major south koreaś rivers using deep learning models, Int. J. Environ. Res.
John Wiley & Sons, Ltd.
Public Health 15 (7) (2018) http://dx.doi.org/10.3390/ijerph15071322, URL
[21] Jingwei Yang, Andreas Holbach, Andre Wilhelms, Yanwen Qin, Binghui Zheng,
https://www.mdpi.com/1660-4601/15/7/1322. Number: 1322 tex.pubmedid:
Hua Zou, Boqiang Qin, Guangwei Zhu, Stefan Norra, Highly time-resolved
29937531.
analysis of seasonal water dynamics and algal kinetics based on in-situ multi-
[39] Zhenyu Yu, Kun Yang, Yi Luo, Chunxue Shang, Spatial-temporal process sim-
sensor-system monitoring data in Lake Taihu, China, Sci. Total Environ. 660
ulation and prediction of chlorophyll-a concentration in Dianchi Lake based
(2019) 329–339, http://dx.doi.org/10.1016/j.scitotenv.2019.01.044, Publisher:
on wavelet analysis and long-short term memory network, J. Hydrol. 582
Elsevier B.V..
(2020) 124488, http://dx.doi.org/10.1016/j.jhydrol.2019.124488, URL https://
[22] Mengyuan Zhu, Jiawei Wang, Xiao Yang, Yu Zhang, Linyu Zhang, Hongqiang
www.sciencedirect.com/science/article/pii/S0022169419312235.
Ren, Bing Wu, Lin Ye, A review of the application of machine learning in water
quality evaluation, Eco-Environ. Health (2022) http://dx.doi.org/10.1016/j.eehl. [40] Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, Yoshua Bengio, N-BEATS:
2022.06.001. Neural basis expansion analysis for interpretable time series forecasting, in:
[23] Rahim Barzegar, Mohammad Taghi Aalami, Jan Adamowski, Short-term water Eighth International Conference on Learning Representations, 2020.
quality variable prediction using a hybrid CNN–LSTM deep learning model, [41] Boris N. Oreshkin, Grzegorz Dudek, Paweł Pełka, Ekaterina Turkina, N-BEATS
Stoch. Environ. Res. Risk Assess. 34 (2) (2020) 415–433, http://dx.doi.org/ neural network for mid-term electricity load forecasting, Appl. Energy 293 (2021)
10.1007/S00477-020-01776-2, URL https://link.springer.com/article/10.1007/ 116918, http://dx.doi.org/10.1016/j.apenergy.2021.116918.
s00477-020-01776-2. Publisher: Springer. [42] B. Puszkarski, K. Hryniów, G. Sarwas, Comparison of neural basis expansion
[24] Hu Li, Chengxin Qin, Weiqi He, Fu Sun, Pengfei Du, Improved predictive analysis for interpretable time series (N-BEATS) and recurrent neural networks
performance of cyanobacterial blooms using a hybrid statistical and deep- for heart dysfunction classification, Physiol. Meas. 43 (6) (2022) 064006, http:
learning method, Environ. Res. Lett. 16 (12) (2021) http://dx.doi.org/10.1088/ //dx.doi.org/10.1088/1361-6579/ac6e55.
1748-9326/ac302d. [43] Thierry Rock Jossou, Zakaria Tahori, Godwin Houdji, Daton Medenou, Abdelali
[25] Alberto Mozo, Jesús Morón-López, Stanislav Vakaruk, Ángel G. Pompa-Pernía, Lasfar, Fréjus Sanya, Mêtowanou Héribert Ahouandjinou, Silvio M. Pagliara,
Ángel González-Prieto, Juan Antonio Pascual-Aguilar, Sandra Gómez-Canaval, Muhammad Salman Haleem, Aziz Et-Tahir, N-beats as an EHG signal forecast-
Juan Manuel Ortiz, Chlorophyll soft - sensor based on machine learning models ing method for labour prediction in full term pregnancy, Electronicsweek 11
for algal bloom predictions, Sci. Rep. (2022) 1–23, http://dx.doi.org/10.1038/ (22) (2022) 3739, http://dx.doi.org/10.3390/electronics11223739, URL https:
s41598-022-17299-5, ISBN: 0123456789 Publisher: Nature Publishing Group UK. //www.mdpi.com/2079-9292/11/22/3739.

18
S. Martín-Suazo et al. Knowledge-Based Systems 301 (2024) 112279

[44] Lijie Deng, Ke Ruan, Xun Chen, Xiaoying Huang, Yongqing Zhu, Weihao Yu, [53] Musaed Alhussein, Khursheed Aurangzeb, Syed Irtaza Haider, Hybrid CNN-LSTM
An IP network traffic prediction method based on ARIMA and N-BEATS, in: Model for Short-Term Individual Household Load Forecasting, IEEE Access :
2022 IEEE 4th International Conference on Power, Intelligent Computing and Pract. Innov. Open Solut. 8 (2020) 180544–180557, http://dx.doi.org/10.1109/
Systems, ICPICS, 2022, pp. 336–341, http://dx.doi.org/10.1109/ICPICS55264. ACCESS.2020.3028281.
2022.9873564, tex.eventtitle: 2022 IEEE 4th international conference on power, [54] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (8)
intelligent computing and systems (ICPICS). (1997) 1735–1780.
[45] Stanislav Vakaruk, Amit Karamchandani, Jesús Enrique Sierra-García, Alberto [55] Bryan Lim, Stefan Zohren, Time-series forecasting with deep learning: A survey,
Mozo, Sandra Gómez-Canaval, Antonio Pastor, Transformers for multi-horizon Phil. Trans. R. Soc. A 379 (2194) (2021) 20200209, http://dx.doi.org/10.1098/
forecasting in an industry 4.0 use case, Sensors 23 (7) (2023) 3516, http: rsta.2020.0209, Publisher: Royal Society.
//dx.doi.org/10.3390/s23073516, Publisher: Multidisciplinary Digital Publishing [56] Y. LeCun, B. Boser, J.S. Denker, D. Henderson, R.E. Howard, W. Hubbard, L.D.
Institute. Jackel, Backpropagation applied to handwritten zip code recognition, Neural
[46] Souhaib Ben Taieb, Antti Sorjamaa, Gianluca Bontempi, Multiple-output mod- Comput. 4 (1) (1989) 541–551.
eling for multi-step-ahead time series forecasting, Neurocomputing 73 (10–12) [57] K. Fukushima, Neocognitron: A self-organizing neural network model for a
(2010) 1950–1957, http://dx.doi.org/10.1016/j.neucom.2009.11.030. mechanism of pattern recognition unaffected by shift in position, Biol. Cybernet.
[47] Konstantinos Benidis, Syama Sundar Rangapuram, Valentin Flunkert, Yuyang 36 (4) (1980) 193–202.
Wang, Danielle Maddix, Caner Turkmen, Jan Gasthaus, Michael Bohlke- [58] Alberto Mozo, Bruno Ordozgoiti, Sandra Gomez-Canaval, Forecasting short-term
Schneider, David Salinas, Lorenzo Stella, François-Xavier Aubet, Laurent Callot, data center network traffic load with convolutional neural networks, PLoS One
Tim Januschowski, Deep learning for time series forecasting: Tutorial and 13 (2) (2018) e0191939.
literature survey, ACM Comput. Surv. 55 (6) (2022) 121:1–121:36, http://dx. [59] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, Kilian Q. Weinberger,
doi.org/10.1145/3533382. Densely Connected Convolutional Networks, IEEE Computer Society, 2017,
[48] Fatoumata Dama, Christine Sinoquet, Time series analysis and modeling to pp. 2261–2269, http://dx.doi.org/10.1109/CVPR.2017.243, URL https://www.
forecast: A survey, 2021, arXiv:2104.00164 [cs].. computer.org/csdl/proceedings-article/cvpr/2017/0457c261/12OmNBDQbld.
[49] Sima Siami-Namini, Neda Tavakoli, Akbar Siami Namin, A comparison of ARIMA tex.eventtitle: 2017 IEEE conference on computer vision and pattern recognition
and LSTM in forecasting time series, in: 2018 17th IEEE International Conference (CVPR).
on Machine Learning and Applications, ICMLA, 2018, pp. 1394–1401, http: [60] Amit Karamchandani, Alberto Mozo, Stanislav Vakaruk, Sandra Gómez-Canaval,
//dx.doi.org/10.1109/ICMLA.2018.00227. J Enrique Sierra-García, Antonio Pastor, Using N-BEATS ensembles to pre-
[50] Hansika Hewamalage, Christoph Bergmeir, Kasun Bandara, Recurrent neural dict automated guided vehicle deviation, Appl. Intell. (2023) 1–66, Publisher:
networks for time series forecasting: Current status and future directions, Int. Springer.
J. Forecast. 37 (1) (2021) 388–427, http://dx.doi.org/10.1016/j.ijforecast.2020. [61] Ingrid Chorus, Jaime Bartram, Toxic cyanobacteria in water. a guide to their
06.008. public health consequences, monitoring, and management, in: Limnology and
[51] Taoying Li, Miao Hua, Xu Wu, A Hybrid CNN-LSTM Model for Forecasting Oceanography, Vol. 45, 1999, p. 1212, http://dx.doi.org/10.4319/lo.2000.45.5.
Particulate Matter (PM2.5), IEEE Access : Pract. Innov. Open Solut. 8 (2020) 1212, ISSN: 00243590 Number: 5.
26933–26940, http://dx.doi.org/10.1109/ACCESS.2020.2971348. [62] Peter Hansen, James Nason, Asger Lunde, The model confidence set,
[52] Xifeng Guo, Qiannan Zhao, Di Zheng, Yi Ning, Ye Gao, A short-term load Econometrica 79 (2010) 453–497, http://dx.doi.org/10.2139/ssrn.522382.
forecasting model of multi-scale CNN-LSTM hybrid neural network considering
the real-time electricity price, Energy Rep. 6 (2020) 1046–1053, http://dx.doi.
org/10.1016/j.egyr.2020.11.078.

19