sensors

Article
CNN-LSTM vs. LSTM-CNN to Predict Power Flow Direction:
A Case Study of the High-Voltage Subnet of Northeast Germany
Fachrizal Aksan 1 , Yang Li 2 , Vishnu Suresh 1, * and Przemysław Janik 1

1 Faculty of Electrical Engineering, Wroclaw University of Science and Technology, 50-370 Wroclaw, Poland;
[email protected] (F.A.); [email protected] (P.J.)
2 Department of Energy Distribution and High Voltage Engineering, Brandenburg University of Technology
Cottbus-Senftenberg, 03046 Cottbus, Germany; [email protected]
* Correspondence: [email protected]

Abstract: The massive installation of renewable energy sources together with energy storage in the
power grid can lead to fluctuating energy consumption when there is a bi-directional power flow
due to the surplus of electricity generation. To ensure the security and reliability of the power grid,
high-quality bi-directional power flow prediction is required. However, predicting bi-directional
power flow remains a challenge due to the ever-changing characteristics of power flow and the
influence of weather on renewable power generation. To overcome these challenges, we present two
of the most popular hybrid deep learning (HDL) models based on a combination of a convolutional
neural network (CNN) and long-term memory (LSTM) to predict the power flow in the investigated
network cluster. In our approach, the models CNN-LSTM and LSTM-CNN were trained with two
different datasets in terms of size and included parameters. The aim was to see whether the size of
the dataset and the additional weather data can affect the performance of the proposed model to
predict power flow. The result shows that both proposed models can achieve a small error under
certain conditions. While the size and parameters of the dataset can affect the training time and
accuracy of the HDL model.

Keywords: CNN-LSTM; LSTM-CNN; power flow prediction; network cluster

Citation: Aksan, F.; Li, Y.; Suresh, V.;
Janik, P. CNN-LSTM vs. LSTM-CNN
to Predict Power Flow Direction:
A Case Study of the High-Voltage 1. Introduction
Subnet of Northeast Germany.
Energy is the most important element for global economic growth [1]. However, to
Sensors 2023, 23, 901. https://
meet the needs of the global market, mainly fossil fuels are used. With the depletion of fossil
doi.org/10.3390/s23020901
fuel supplies and the increasing demand for electrical energy, we are now massively turning
Academic Editor: Alessandro to renewable energy sources as an alternative to provide a more economical and diversified
Mingotti energy mix that ensures energy security and sustainability [2]. The enormous expansion
and installation of renewable energy sources, such as wind turbines and photovoltaics
Received: 24 November 2022
Revised: 7 January 2023
(PV), in electrical grids can lead to a change in the power flow and also affect the extent
Accepted: 9 January 2023
of the bi-directional power flow via a transformer to or from the power grid system [3,4].
Published: 12 January 2023 Furthermore, this phenomenon has also led to a significant change in the traditional
power supply mechanisms, from a centralised to a decentralised power system and from a
directional to a bi-directional power supply system [4]. In the past, the electricity supply
mechanism operated by generating electricity in large central power plants on the high-
Copyright: © 2023 by the authors. voltage grid, which was then transmitted to end users through the transmission and
Licensee MDPI, Basel, Switzerland. distribution system. Therefore, power flows on the generation and consumption side were
This article is an open access article much more predictable. This is due to the advance planning of power plant deployment
distributed under the terms and and the standardized creation of load profiles [3]. However, most renewables work in
conditions of the Creative Commons the opposite way, they can be installed together with energy storage at all grid levels, but
Attribution (CC BY) license (https:// mainly at the medium and lower voltage levels of the distribution grid system, so that
creativecommons.org/licenses/by/
integration becomes easier.
4.0/).

Sensors 2023, 23, 901. https://doi.org/10.3390/s23020901 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 901 2 of 20

Therefore, the simple and decentralised integration of energy resources into the power
system can bring new challenges. These include the challenge of predicting power flow
due to the intermittent nature of renewables and also the occurrence of bi-directional power
flows in the power system [5]. Bi-directional power flows have two opposite directions like
tides. They can occur, for example, when renewable electricity generators become common
for domestic use and a surplus of electricity generated by this generator is passed on to an
electricity utility. In this case, the household electricity grid system will have a bi-directional
power flow. Another challenge is the difficulty of predicting consumption behaviour with
standardised load profiles. This is due to the numerous integrations of energy storage and
e-mobility into the power grid [3], so that several end users have different types of loads,
which are mostly inductive or capacitive.
Knowing the power flow in the grid is very important to know how much electricity
is imported when the regional load is high and how much surplus power generation is
exported from the investigated region. This can also help the distribution utilities to control
the generated electricity that is transmitted to the load centres and distributed from there to
the end-users. Furthermore, it can also help grid operators avoid congestion and manage
the exchange of energy between the different connected electricity suppliers. Knowledge of
the actual conditions and power flow on the grid is indeed important, both in the present
and future state. However, all the above challenges in integrating renewable energy have
led to unpredictable power flows in the grid, especially at transformers between voltage
levels. Therefore, this study investigates the prediction of the bi-direction of power flow to
meet these grid operation strategies.
There are several ways to make an accurate prediction. Currently, the development
of machine learning (ML) and deep learning (DL) is increasing with the advances in the
field of computer science. Therefore, it is not surprising that the methods using ML and DL
are already used for all prediction and forecasting tasks in modern power systems [6–9].
The competitiveness of DL as a robust method capable of solving complex problems with
a lower error rate has been demonstrated in reference [10]. In general, deep learning can
work as a single network or as a hybrid network. A study in reference [11] examined
several research works on load forecasting using different techniques, and one of the results
shows that the hybrid method has higher accuracy compared to single model algorithms.
Apart from the advantages, the DL method still has the challenge of finding the optimal
configuration for the structure of the models.
To predict the direction of power flow using the DL method, the DL method must
learn from historical power flow measurement data. However, the power flow can also
be influenced by grid topologies and fluctuation generation, as renewable energies are
dependent on the weather condition. Therefore, it is necessary to include weather data in
the prediction of bi-directional power flow. In this paper, we present an approach to predict
bi-directional power flow that takes into account power measurement data and weather
data. The proposed prediction model is based on a hybrid deep learning (HDL) model using
a convolutional neural network (CNN) and a long-term memory (LSTM) as the backbone.
An LSTM is a special model that is usually used for time series predictions [12–17], while a
CNN network is mainly used for processing images. However, this model is still suitable
for time series prediction [18–21]. To answer the question of whether the HDL model is
suitable for predicting the direction of power flow, we compare the performance of the
CNN-LSTM model with the LSTM-CNN model to predict the direction of power flow in
the single feeder of the power grid under study. In general, the contribution of our paper is
summarized as follows:
• We present and analyse a hybrid deep learning model for predicting the direction of
power flow in a single feeder of the power grid.
• We compare and evaluate the performance of the proposed HDL model (CNN-LSTM
versus LSTM-CNN) with baseline models (CNN only and LSTM only) for power
flow prediction.
Sensors 2023, 23, 901 3 of 20

• We investigate how weather data can influence the prediction result to determine the
direction of power flow in the studied power system.
• We investigate how the HDL model behaves when the input data used has a different
size and different parameters.
The rest of the paper is organised as follows: related work is presented in Section 2, a
brief description of deep learning model structures is given in Section 3, Section 4 gives a
basic description of the proposed methodology, Section 5 describes our case study and the
dataset used, Section 6 presents the results and discussion, and the last section summarises
the conclusions of this paper.

2. Related Work
Few scientific papers have been published in the field of bi-directional power flow
prediction. However, there are some research papers dealing with power flow prediction
that have similarities with our work. In reference [3], the authors used an LSTM network
for vertical power flow forecasting at a transformer located between the medium- and
high-voltage networks. The authors used an updating process where models that were
trained regularly were checked. The experimental results showed that the proposed
approach achieved a good improvement compared to the approach without an updating
process. The study in reference [22] presents adaptive power flow prediction. The authors
used a machine learning model to predict the voltages of the terminal nodes on the high-
and low-voltage sides of the distribution network without knowing the line topology
or impedance. The proposed model was comparable to the conventional impedance
estimation from power flow analysis. In the study [23], the prediction of power flow was
carried out using an artificial neural network (ANN). The main objective of the authors
was to reduce the maximum prediction error of the model in predicting the magnitude of
the bus voltage and the line load by applying appropriate data pre-processing techniques.
A power flow analysis can be performed to learn more about the bi-directional power flow
in an integrated power system with other power resources. Knowledge of bi-directional
power flow is very important for power system operators to plan, maintain and modify
circuits for facilities or loads that may require power during peak load or feed surplus
power to the main power supply. In our study case, we performed two HDL models:
CNN-LSTM and LSTM-CNN, to predict the direction of power flow on individual feeder
lines of the power grid under study.
CNN-LSTM and LSTM-CNN are well-known models used to solve prediction and
forecasting tasks in the field of power systems. For example, in reference [24], the proposed
model CNN-LSTM was used to predict the hourly heating load. The authors employed
the CNN to extract the spatial features and influencing factors of the heating load data,
while the LSTM was used as a temporal feature extractor to extract the time lag features of
the heating load data. Therefore, the proposed model is more suitable for predicting the
heating load in the presence of nonlinearity and significant thermal inertia delay. Another
study using the CNN-LSTM architecture can be found in reference [25]. In this work, a
hybrid model was proposed to predict the short-term photovoltaic electricity generation.
Based on the simulation results, it was shown that CNN-LSTM has excellent performance
in terms of stability, accuracy and prediction compared to the standard algorithm ML and
the single model DL.
If CNN-LSTM is proven to have good capabilities, then so is LSTM-CNN. A study
in reference [26] proved that LSTM-CNN can achieve excellent results in load forecasting.
The authors used the proposed model to predict the load of the next time step on different
datasets. The result showed that the proposed model still has high accuracy. In a similar
case of load prediction, LSTM-CNN was also used in reference [27]. In this study, LSTM
was used to increase the sensitivity of the model and enhance the influence of important
information in the features, while CNN was added to improve the model’s ability to
perceive the sensitivity to data. In this way, the model achieved an excellent predictive
Sensors 2023, 23, 901 4 of 20

performance. The summary of the research that deals with power flow and the evidence
that CNN-LSTM and LSTM-CNN are good for prediction is presented in Table 1.

Table 1. Summary of related work.

Topic Reference Methodology/Output

Brauns et al. [3] Using LSTM model for vertical power flow forecasting.

Prediction on Park et al. [22] Implementing machine learning for predicting adaptive power flow.
power flow Schafer et al. [23] Proposed ANN with data pre-processing techniques to solve power flow prediction.
Our Proposed CNN-LSTM and LSTM-CNN models for bi-directional power flow prediction.
Song et al. [24] CNN-LSTM performs excellently in the area of predictive accuracy.
CNN-LSTM and Agga et al. [25] CNN-LSTM performs better than the standard ML or individual DL.
LSTM-CNN for
prediction Farsi et al. [26] Parallel LSTM-CNN is a good candidate for use as a short-term prediction tool.
Li et al. [27] Parallel LSTM-CNN has a good prediction effect.

3. Deep Learning Model Structures

In this section, we present a brief description of the proposed HDL model for predicting
the direction of power flow in a single feeder of the studied network cluster. Since the
CNN-LSTM and LSTM-CNN models have been proposed in this paper. It is necessary
to briefly discuss the LSTM and CNN networks. The reason is that these two individual
models form the backbone of the proposed HDL model.

3.1. Long Short-Term Memory (LSTM) Network

The LSTM is a special type of recurrent neural network (RNN) first introduced by
Hochreiter and Schmidhuber [28]. This model was developed to circumvent the problem
of long-term dependence and to solve the vanishing gradient problem, since the standard
RNN model is not able to learn long-term dependence. Therefore, this model includes
memory cells and gates to regulate the network’s information and remember information
over long periods of time. Indeed, the LSTM is a popular DL model used for all cases
of forecasting and prediction. In a study in reference [29], the LSTM network showed
good performance in predicting network losses in a Finnish power grid. Moreover, it can
perform very well even without prior knowledge of the power system. Another study
using LSTM can be found in reference [30]. The authors implement the proposed LSTM
model to forecast hourly day-ahead solar irradiance based on weather forecast data. The
proposed model performs slightly better than some competing algorithms because it takes
into account the dependencies between different hours of the same day.
In general, an LSTM network consists of memory blocks called cells. Each cell has
two states: the cell state and the hidden state. The cells in the LSTM network are used to
make important decisions by storing or ignoring information about important components.
These components are called gates, which are organised as follows: forget gates, input
gates, and output gates. According to the structures shown in Figure 1, the LSTM model
operates in three stages: In the first stage, the network works with the forget gate to check
what kind of information needs to be ignored or stored for the cell state. The calculation
starts by considering the input at the current time step (xt ) and the previous value of the
hidden state (h(t−1) ) using the sigmoid function (S). The formula for the calculation in forget
gate is as follows.  h i 
f t = S w f · h ( t −1) , x t + b f (1)
 memory, containing information about previous data and used for predictions. To deter‐
mine the value of the hidden state, the calculation must have the reference value of the 
new cell state and the output gate (ot). The formula for this process is shown below. 

𝑜𝑡 𝑆 𝑤𝑜 ∙ ℎ 𝑡 1 , 𝑥𝑡 𝑏𝑜   (5) 
Sensors 2023, 23, 901 5 of 20
ℎ 𝑜 ∙𝑇 𝐶   (6) 

 
Figure 1. The architecture of LSTM network. 
Figure 1. The architecture of LSTM network.

3.2. Convolutional Neural Network (CNN) 
In the second phase, the calculation of the network continues by converting the old
cell state (C(t−1) ) into a new cell state (Ct ). This process selects which new information must
Another variant of the deep learning model is the CNN network. This model has the 
be included in the long-term memory (cell state). To obtain the new cell state value, the
ability to learn highly abstracted features of objects [31]. Therefore, it is very suitable for 
calculation process should take into account the reference value from the forgetting gate,
visual image analysis and recognition [6,32,33]. However, the CNN model also has a layer 
the input gate and the cell update gate value. The formulas for this step are shown below.
that is able to learn the features of sequence data with multiple variables. Therefore, it can 
also be used for any prediction task. In the case of fault detection, reference [34] shows 
 h i 
it = S wi · h(t−1) , xt + bi
that CNN has been used with attentive density to detect and identify the fault types and  (2)
severities of rolling bearings. In [35], the CNN architecture was used to solve the problem 
 h i 
C 0 t = T wc · h(t−1) , xt + bc
of blade icing by analysing the imbalance of the supervisory control and data acquisition  (3)
(SCADA) data of a wind turbine.   
Ct = C(t−1) · f t + it · C 0 t


According to reference [19], a typical CNN model, as shown in  Figure  2, comprises 
(4)
several layers: convolutional layer, pooling layer, a flattening layer, and a fully connected 
Once the cell status update is complete, the final step is to determine the value
layer. The convolutional layer is the main component of the CNN network, which oper‐
of the
ates  on hidden state (hof 
the  principle  (t) ).sliding 
The aim of thisand 
windows  process is for
weight  the hidden
sharing  statecomputational 
to  reduce  to act as the
network’s memory, containing information about previous data and used for predictions.
To determine the value of the hidden state, the calculation must have the reference value of
the new cell state and the output gate (ot ). The formula for this process is shown below.
   h i 
ot = S wo · h(t−1) , xt + bo (5)

ht = ot · T (Ct ) (6)

3.2. Convolutional Neural Network (CNN)

Another variant of the deep learning model is the CNN network. This model has the
ability to learn highly abstracted features of objects [31]. Therefore, it is very suitable for
visual image analysis and recognition [6,32,33]. However, the CNN model also has a layer
that is able to learn the features of sequence data with multiple variables. Therefore, it can
also be used for any prediction task. In the case of fault detection, reference [34] shows
that CNN has been used with attentive density to detect and identify the fault types and
severities of rolling bearings. In [35], the CNN architecture was used to solve the problem
of blade icing by analysing the imbalance of the supervisory control and data acquisition
(SCADA) data of a wind turbine.
According to reference [19], a typical CNN model, as shown in Figure 2, comprises
several layers: convolutional layer, pooling layer, a flattening layer, and a fully connected
layer. The convolutional layer is the main component of the CNN network, which operates
on the principle of sliding windows and weight sharing to reduce computational complexity.
In this layer, the kernel method is used to extract various features from the input data.
The next layer is the pooling layer. This layer is designed to reduce the size of the feature
Sensors 2023, 22, x FOR PEER REVIEW  6 of 22 
 

Sensors 2023, 23, 901 6 of 20

complexity. In this layer, the kernel method is used to extract various features from the 
input data. The next layer is the pooling layer. This layer is designed to reduce the size of 
the feature map involved by reducing the connections between layers and running each 
map involved by reducing the connections between layers and running each feature map
feature map independently. The main goal of the pooling operation is to reduce the di‐
independently. The main goal of the pooling operation is to reduce the dimensionality
mensionality  and  extract  the  dominant  features  for  efficient  training  of  the  model  [6]. 
and extract the dominant features for efficient training of the model [6]. There are several
There are several types of pooling operations: max pooling and average pooling. Before 
types of pooling operations: max pooling and average pooling. Before proceeding with
proceeding with the fully connected linked layer (FC), it is necessary to use the flattening 
the fully connected linked layer (FC), it is necessary to use the flattening layer to create a
layer to create a one‐dimensional vector, because the FC layer consists of the weights and 
one-dimensional vector, because the FC layer consists of the weights and biases along with
biases along with the neurons to connect the neurons between the different layers. The FC 
the neurons to connect the neurons between the different layers. The FC layer is sometimes
layer is sometimes inserted as the last layer before the output layer of the CNN network. 
inserted as the last layer before the output layer of the CNN network.

 
Figure 2. The structure of CNN network.
Figure 2. The structure of CNN network. 
3.3. Hybrid Deep Learning Model
3.3. Hybrid Deep Learning Model 
As previously stated, two types of HDL were used for training directional power flow
As  previously 
prediction. They arestated,  two  types 
the models of  HDL and
CNN-LSTM were  used  for  training 
LSTM-CNN. directional 
The structure power 
of the two
flow prediction. They are the models CNN‐LSTM and LSTM‐CNN. The structure of the 
HDL models used in this paper is shown in the following Table 2. The architecture of
two HDL models used in this paper is shown in the following Table 2. The architecture of 
CNN-LSTM (see Figure 3) was developed with CNN layers on the front end. The aim is to
CNN‐LSTM (see 
extract the featuresFigure  3) was developed with CNN layers on the front end. The aim is to 
of the input dataset. The outputs of the CNN layers were then passed
extract the features of the input dataset. The outputs of the CNN layers were then passed 
to the LSTM layers and a dense layer at the output to support sequence prediction.
to the LSTM layers and a dense layer at the output to support sequence prediction. 
Table 2. Hybrid deep learning model structure.
Table 2. Hybrid deep learning model structure. 
HDL Model Structure of Layer
HDL Model  Structure of Layer 
LSTM layer (neurons: 30) + LSTM layer (neurons: 15) + dense layer (neuron: 1,
LSTM LSTM layer (neurons: 30) + LSTM layer (neurons: 15) + dense layer (neu‐
LSTM  LeakyRelu activation)
ron: 1, LeakyRelu activation) 
conv1D layer (filters: 16, filter size: 3, relu activation) + conv1D layer (filters: 16,
CNN conv1D layer (filters: 16, filter size: 3, relu activation) + conv1D layer (fil‐
filter size: 3, relu activation) + maxpooling1D (polling size: 2, padding: same) +
ters: 16, filter size: 3, relu activation) + maxpooling1D (polling size: 2, 
flatten layer+ dense layer (neuron: 1, LeakyRelu activation)
CNN 
padding: same) + flatten layer+ dense layer (neuron: 1, LeakyRelu acti‐
Conv1D layer (filters: 32, filter size: 3, relu activation) + conv1D layer (filters: 32,
vation) 
filter size: 3, relu activation) + maxpooling1D (polling size: 2, padding:same) +
CNN-LSTM
flatten layer + LSTM layer (neurons: 32, relu activation) + LSTM layer (neurons:
Conv1D layer (filters: 32, filter size: 3, relu activation) + conv1D layer 
10) + dense layer (neuron: 1, LeakyRelu activation)
(filters: 32, filter size: 3, relu activation) + maxpooling1D (polling size: 2, 
CNN‐LSTM  LSTM layer (neurons: 32, relu activation) + LSTM layer (neurons: 10) + conv1D
padding:same) + flatten layer + LSTM layer (neurons: 32, relu activation) 
layer (filters: 32, filter size: 3, relu activation) + conv1D layer (filters: 32, filter
Sensors 2023, 22, x FOR PEER REVIEW  LSTM-CNN + LSTM layer (neurons: 10) + dense layer (neuron: 1, LeakyRelu activa‐ 7 of 22 
  size: 3, relu activation) + maxpooling1D (polling size: 2, padding: same) + flatten
tion) 
layer+ dense layer (neuron: 1, LeakyRelu activation)
LSTM layer (neurons: 32, relu activation) + LSTM layer (neurons: 10) + 
conv1D layer (filters: 32, filter size: 3, relu activation) + conv1D layer (fil‐
LSTM‐CNN  ters: 32, filter size: 3, relu activation) + maxpooling1D (polling size: 2, 
padding: same) + flatten layer+ dense layer (neuron: 1, LeakyRelu acti‐
vation) 

 
Figure 3. The structure of CNN-LSTM model.
Figure 3. The structure of CNN‐LSTM model. 

On
On the
the other hand,
other  the
hand,  structure
the  of LSTM-CNN
structure  (see (see 
of  LSTM‐CNN  Figure 4) is 4) 
Figure  in ais different sequence.
in  a  different  se‐
The LSTM layers were used to order the sequence of time series data as input. The idea
quence. The LSTM layers were used to order the sequence of time series data as input. 
behind this is that the output of the LSTM layers contains more new information, which
The idea behind this is that the output of the LSTM layers contains more new information, 
 
which is then fed into the CNN layers to extract local features. The output of this convo‐
lutional layer is then pooled into a smaller dimension and passed to the dense layer to 
predict the final output. 
  
Figure 3. The structure of CNN‐LSTM model. 

Sensors 2023, 23, 901 7 ofse‐

On  the  other  hand,  the  structure  of  LSTM‐CNN  (see  Figure  4)  is  in  a  different  20
quence. The LSTM layers were used to order the sequence of time series data as input. 
The idea behind this is that the output of the LSTM layers contains more new information, 
which is then fed into the CNN layers to extract local features. The output of this convo‐
is then fed into the CNN layers to extract local features. The output of this convolutional
lutional layer is then pooled into a smaller dimension and passed to the dense layer to 
layer is then pooled into a smaller dimension and passed to the dense layer to predict the
final output.
predict the final output. 

Sensors 2023, 22, x FOR PEER REVIEW  8 of 22 
 

error. Therefore, to avoid this problem, it was necessary to scale or normalise all variables 
in the dataset. Moreover, this can also improve the performance of HDL as all input vari‐
ables are scaled to a standard range [37]. In this study, the numerical scaling method min–
 
max normalisation was used. The formula for converting the original value into a normal‐
Figure 4. The structure of LSTM-CNN model.
Figure 4. The structure of LSTM‐CNN model. 
ised value is shown in the following equation. 
4. Methodology
4. Methodology  𝑥 𝑚𝑖𝑛 𝑥
To compare the performance of𝑥CNN-LSTM and LSTM-CNN   (7) 
in predicting the direction
𝑚𝑎𝑥 𝑥 𝑚𝑖𝑛 𝑥
To compare the performance of CNN‐LSTM and LSTM‐CNN in predicting the direc‐
of load flow of each line in the power grid, the proposed methodology is shown in Figure 5.
tion of load flow of each line in the power grid, the proposed methodology is shown in 
where x′ is the normalised value, x is the original value, max(x) is the maximum value of 
The proposed approach is divided into four steps: data collection, pre-processing of the data,
Figure 5. The proposed approach is divided into four steps: data collection, pre‐processing 
x and min(x) is the minimum value of x. 
creation of the models CNN-LSTM and LSTM-CNN and evaluation of the models.
of  the  data,  creation  of  the  models  CNN‐LSTM  and  LSTM‐CNN  and  evaluation  of  the 
models. 

4.1. Step 1: Data Collection 
Data collection is a crucial step because all further steps depend on the availability of 
the  data.  Data  collection  is  about  gathering  all  the  necessary  data  from  the  available 
sources. It is important to clean and filter the data before it is used. In this work, the raw 
data  for  directional  power  flow  is  collected  from  a  power  grid  under  study  and  the 
weather data is obtained from a weather service provider. Therefore, the quality of the 
data used in this work is suitable for training and testing the proposed hybrid deep learn‐
ing model for predicting the direction of power flow. 

4.2. Step 2: Data Pre‐processing 
After data collection, the next step is to pre‐process the data. The main objective of 
this phase is to prepare and convert the raw data into a format suitable for the HDL model. 
The implementation of data pre‐processing is very important for any type of deep learn‐
ing model as it can improve the model accuracy by improving the quality of the data and 
extracting  valuable  information  from  the  data  [36].  In  this  work,  various  data  pre‐pro‐
cessing techniques were used, ranging from normalising the data to splitting the dataset. 

4.2.1. Data Normalization 
The datasets used in this study come from different sources, and their parameters 
have different units and scales. These differences in the datasets may affect the perfor‐
mance of HDL during the learning process and, even worse, increase the generalisation 

 
Figure 5. Proposed methodology. 
Figure 5. Proposed methodology.

4.2.2. Dataset Splitting 
4.1. Step 1: Data Collection
For the development and evaluation of a predictive model. Sometimes the input data 
Data collection is a crucial step because all further steps depend on the availability
must be prepared in a suitable way and divided into a training, a validation and a test 
of the data. Data collection is about gathering all the necessary data from the available
dataset. In principle, there is no optimal percentage for the splitting ratio. However, there 
are several ways to split the dataset, e.g., 90% for training and 10% for testing [38,39], or 
80% for training and 20% for testing [40,41]. However, this study refers to the scenario of 
70% for the training dataset, 15% for the validation dataset and 15% for the test dataset, 
based  on  references  [15,42–45].  The  training  and  validation  dataset  was  split  using  the 
Sensors 2023, 23, 901 8 of 20

sources. It is important to clean and filter the data before it is used. In this work, the raw
data for directional power flow is collected from a power grid under study and the weather
data is obtained from a weather service provider. Therefore, the quality of the data used in
this work is suitable for training and testing the proposed hybrid deep learning model for
predicting the direction of power flow.

4.2. Step 2: Data Pre-Processing

After data collection, the next step is to pre-process the data. The main objective
of this phase is to prepare and convert the raw data into a format suitable for the HDL
model. The implementation of data pre-processing is very important for any type of
deep learning model as it can improve the model accuracy by improving the quality of
the data and extracting valuable information from the data [36]. In this work, various
data pre-processing techniques were used, ranging from normalising the data to splitting
the ataset.

4.2.1. Data Normalization

The datasets used in this study come from different sources, and their parameters have
different units and scales. These differences in the datasets may affect the performance
of HDL during the learning process and, even worse, increase the generalisation error.
Therefore, to avoid this problem, it was necessary to scale or normalise all variables in the
dataset. Moreover, this can also improve the performance of HDL as all input variables
are scaled to a standard range [37]. In this study, the numerical scaling method min–max
normalisation was used. The formula for converting the original value into a normalised
value is shown in the following equation.

x − min( x )
x0 = (7)
max ( x ) − min( x )

where x0 is the normalised value, x is the original value, max(x) is the maximum value of x
and min(x) is the minimum value of x.

4.2.2. Dataset Splitting

For the development and evaluation of a predictive model. Sometimes the input data
must be prepared in a suitable way and divided into a training, a validation and a test
dataset. In principle, there is no optimal percentage for the splitting ratio. However, there
are several ways to split the dataset, e.g., 90% for training and 10% for testing [38,39], or
80% for training and 20% for testing [40,41]. However, this study refers to the scenario of
70% for the training dataset, 15% for the validation dataset and 15% for the test dataset,
based on references [15,42–45]. The training and validation dataset was split using the train
test split library of the scikit learn framework. In splitting the data, we performed a control
shuffle of the data with a random state value of 42.
Before dividing the dataset, we categorised two datasets by size and parameters in the
proposed methodology. The aim of this approach was to follow our research contribution
to find out the extent to which weather data can influence the proposed HDL model to
predict the direction of power flow at each line in the power grid under study. Therefore,
the first group of datasets contained only directional power flow data, while the dataset of
the second group contained directional power flow and weather data.

4.3. Step 3: Build Prediction Models

Our proposed work focuses on predicting the direction of power flow in each line
of the power system under study using two types of HDL models. The first model is the
CNN-LSTM and the second model is the LSTM-CNN. Both HDL models use the same
two individual networks of CNN and LSTM. However, they are just constructed in a
different order. When building an HDL model, there is no direct information about the
optimal model architecture for a particular model. For example, how many hidden layers,
Sensors 2023, 23, 901 9 of 20

activation functions and optimisers should be used. Therefore, it is necessary to select

an optimal set of parameter configurations, as these parameters are used to control and
manage the learning process of the HDL model to accurately predict the output. There
are different ways to tune the hyperparameters. In this study, we randomly selected the
parameters to tune the hyperparameters of the proposed HDL models and the baseline
models (CNN only and LSTM only).
In this work, the CNN-LSTM architecture integrated several layers. In the first layer of
the model, there are two stacks of convolutional layers with the activation function rectified
linear unit (relu). This is followed by the max-pooling layer and the flattening layer. After
this, the stack LSTM layer with the activation function relu was built. A dense layer is
connected as the output layer. This layer uses the activation function leaky rectified linear
unit (LeakyRelu). As the optimiser, we choose Adam with a learning rate of 0.01 and a loss
function with mean square error.
On the other hand, for the LSTM-CNN architecture. We used the same layers and
activation functions as in the CNN-LSTM model. The only difference is the order of the
layers. The model LSTM-CNN was built from stacks of the LSTM network, which were
then passed to stacks of convolutional layers, followed by a sequence of max-pooling
layers and the flattening layer. In the output there is a dense layer. The optimiser and loss
function implemented in the LSTM-CNN model are the same as in the CNN-LSTM model.
A summary of the architectures of CNN-LSTM, LSTM-CNN and the baseline models can
be found in Table 2.
There are several important hyperparameters that need to be set in the proposed HDL
and baseline models, such as the number of batch sizes and the number of epochs. The
batch size is a parameter that specifies the number of samples that are run before the model
parameters are updated. Epoch number is a hyperparameter that specifies how often the
learning algorithm is applied to the training dataset. Using too few or too many epochs
may result in under-fitting or over-fitting. In this work, for each proposed HDL model
and baseline model during the training process, we set the same batch size with a value of
32 and the same number of epochs with a value of 25. All structures and layers of all models
were created using TensorFlow and the Keras library. During the experimental research,
the proposed HDL along with baseline models were trained and tested on a laptop with
the technical specifications listed in Table 3.

Table 3. Machine specification.

Parameter Specification
CPU Intel(R) Core(TM) i5-7200U CPU @ 2.50 GHz 2.71 GHz
GPU Intel UHD Graphics 620
HDD/SDD 750 GB
RAM 16 GB
OS Windows 10 pro 64-bit

4.4. Step 4: Evaluate the Proposed HDL Model

Model evaluation is a very important step to assess and measure the performance
and accuracy of the proposed HDL model using the metric scores. The evaluation metrics
used for this study were selected based on the recommendations of studies and reports
in the field of predictive cases. The metrics are the root mean square error (RMSE), the
mean absolute error (MAE) and the coefficient of determination (R2 ). The formula for these
metrics is presented in the following equations.
s
2
∑tN=1 (yt − ŷt)
RMSE = (8)
N

∑tN=1 |yt − ŷt|

MAE = (9)
N
Sensors 2023, 23, 901 10 of 20

2
∑t (ŷt − y)
R2 = 2
(10)
∑t (yt − y)
where yt is the actual value, ŷt is the predicted value, y is the mean value of y, and
N is the number of observations. The evaluation metrics presented in the Results and
Discussion section were calculated based on the original data. The original data was
obtained by converting the normalized value with the inverse of the min–max scaling
algorithm presented in Equation (7).
Sensors 2023, 22, x FOR PEER REVIEW  11 of 22 
 
5. Case Study and Dataset Description
The complexity of the electricity system is further increased when large power plants
feeding electricity into the transmission grid integrate with countless decentralized renew-
generation is exported from the grid cluster under study. Based on the measurement of 
able energy plants feeding electricity into the medium- and low-voltage grids. This is
directional power, as shown in Figure 6, the sign of the power flow indicates the direction 
because fluctuating electricity generation from renewable energy sources (RES) and the
of power towards or away from the busbar. In this paper, we investigated two HDL mod‐
varying behaviour of electricity consumers lead to a change in the behaviour of the line
els to predict the direction of power flow on a single line of the network under study by 
power flow. In order to reduce the complexity of the entire power grid system and to
considering other existing lines. For example, to predict line 1 of the studied network clus‐
analyse the effects of decentralized power generation from RES, a regional grid network
ter, we used the other lines (line 2, line 3, line 4, line 5, and line 6) as input references. This 
cluster is needed. For this purpose, a simplified grid network cluster is created at the
implementation also applies to the other lines if they are to be predicted. Since the high 
connection point between transmission system operators (TSO) and distribution system
regional installation of renewable energy systems, about 365 MW photovoltaic systems 
operators (DSO) using the following grid reduction procedure. First, the power grid is
and 630 MW wind turbines have been connected to the DSO grids studied locally. A wide 
zoned according to the high-voltage lines and then the internal connection lines are ne-
range of weather data is also used in this work, as local weather has a major impact on 
glected. The area between two transformer substations forms a power supply area, which
regional  electricity  generation  from  renewable  energy  sources.  Therefore,  to  find  out 
is called a network cluster in this study. Within this network cluster, there are different
whether weather data can influence the power flow prediction results, we included addi‐
voltage levels of loads and electricity suppliers. In this paper, one regional high-voltage
tional weather condition parameters as reference values, together with power flow values 
subnet from northeast Germany was taken into account as an example of a network cluster.
that exist on other feeder lines. 
The structure of the network cluster can be seen in Figure 6.

 
Figure 6. Investigated network cluster.
Figure 6. Investigated network cluster. 

The investigated network cluster is supplied by six feeder lines, of which four feed
5.1. Bi‐Directional Power Flow Measurement Data 
lines (line 3, line 4, line 5, line 6) are connected to substation A (Sub_A) and two feed lines
(line To predict the direction of power flow on an individual feeder line of the investigated 
1, line 2) to substation B (Sub_B). By measuring the power of the feeder lines, we can
network cluster. We used raw directional power measurement data provided by the local 
capture the main generation and load information of the grid cluster, how much power
distribution system operator (DSO). This directional power flow data has a temporal res‐
is imported when the regional load of the grid cluster is high, and how much surplus
olution of 15 min and ranges from 1 January 2019 to 31 December 2019. As shown in Fig‐
generation is exported from the grid cluster under study. Based on the measurement of
ure   7,  a  value 
directional of  active 
power, power 
as shown above 6,
in Figure zero 
themeans 
sign of that  power flow
the power is  flowing  away 
indicates the from  the 
direction
busbar to the cluster, while a value below zero means that power is flowing towards the 
of power towards or away from the busbar. In this paper, we investigated two HDL models
busbar from network cluster. Figure 7 shows an example of the power flow in the network 
to predict the direction of power flow on a single line of the network under study by
cluster studied in January 2019. The statistical description of power flow measurement 
considering other existing lines. For example, to predict line 1 of the studied network
dataset can be found in Table 4. 
cluster, we used the other lines (line 2, line 3, line 4, line 5, and line 6) as input references.
This implementation also applies to the other lines if they are to be predicted. Since the high
regional installation of renewable energy systems, about 365 MW photovoltaic systems and
Sensors 2023, 23, 901 11 of 20

630 MW wind turbines have been connected to the DSO grids studied locally. A wide range
of weather data is also used in this work, as local weather has a major impact on regional
electricity generation from renewable energy sources. Therefore, to find out whether
weather data can influence the power flow prediction results, we included additional  
weather condition parameters as reference
Figure 6. Investigated network cluster.  values, together with power flow values that
exist on other feeder lines.
5.1. Bi‐Directional Power Flow Measurement Data 
5.1. Bi-Directional Power Flow Measurement Data
To predict the direction of power flow on an individual feeder line of the investigated 
To predict the direction of power flow on an individual feeder line of the investigated
network cluster. We used raw directional power measurement data provided by the local 
network cluster. We used raw directional power measurement data provided by the local
distribution system operator (DSO). This directional power flow data has a temporal res‐
distribution system operator (DSO). This directional power flow data has a temporal
olution of 15 min and ranges from 1 January 2019 to 31 December 2019. As shown in Fig‐
resolution of 15 min and ranges from 1 January 2019 to 31 December 2019. As shown in
ure  7,  a  value  of  active  power  above  zero  means  that  power  is  flowing  away  from  the 
Figure 7, a value of active power above zero means that power is flowing away from the
busbar to the cluster, while a value below zero means that power is flowing towards the 
busbar to the cluster, while a value below zero means that power is flowing towards the
busbar from network cluster. Figure 7 shows an example of the power flow in the network 
busbar from network cluster. Figure 7 shows an example of the power flow in the network
cluster studied in January 2019. The statistical description of power flow measurement 
cluster studied in January 2019. The statistical description of power flow measurement
dataset can be found in Table 4. 
dataset can be found in Table 4.

 
Figure 7. Power flow in all lines of network cluster.
Figure 7. Power flow in all lines of network cluster. 

Table 4. Statistical description of the power flow measurement data.

Power Flow Away from Busbar Power Flow to Busbar

  Feed Lines Number of Min Mean Max Number of Min Mean Max
Data Point [MW] [MW] [MW] Data Point [MW] [MW] [MW]
Line 1 3287 2.378 7.472 22.189 29323 −104.809 −41.40 −2.378
Line 2 3434 2.378 7.544 23.559 29073 −104.417 −41.19 −2.378
Line 3 11843 0.878 9.162 34.311 21422 −102.501 −21.030 −1.095
Line 4 10574 0.878 9.536 25.230 22760 −115.122 −25.798 −1.095
Line 5 12377 0.878 5.694 21.919 19951 −74.392 −15.507 −1.095
Line 6 15467 0.878 7.072 25.676 16918 −48.257 −10.656 −1.095

5.2. Weather Data

The weather data used in this paper comes from a German weather service provider.
This provider offers access to the Climate Data Centre (CDC) portal to retrieve weather
data with a temporal resolution of 15 min recorded by various weather measuring stations.
The regional stations are filtered, and the mean weather data is calculated. The parameters
of the weather data are explained in more detail below:
• Ground air temperature (2 m above ground) (◦ C);
• Ground wind speed (10 m above ground) (m/s);
• Solar irradiation (W/m2 ).
The length of the weather data used in this paper is the same as the length of power
measurement data, from 1 January 2019 to 31 December 2019. Figure 8 shows the weather
 tions. The regional stations are filtered, and the mean weather data is calculated. The pa‐
rameters of the weather data are explained in more detail below: 
 Ground air temperature (2 m above ground) (°C); 
 Ground wind speed (10 m above ground) (m/s); 
 Solar irradiation (W/m2). 
Sensors 2023, 23, 901 12 of 20
The length of the weather data used in this paper is the same as the length of power 
measurement data, from 1 January 2019 to 31 December 2019. Figure 8 shows the weather 
conditions in the vicinity of the studied grid cluster in January 2019. The statistical de‐
conditions in the vicinity of the studied grid cluster in January 2019. The statistical descrip-
scription of the weather dataset used in this paper can be found in Table 5. 
tion of the weather dataset used in this paper can be found in Table 5.

 
Figure 8. Local weather data sample.
Figure 8. Local weather data sample. 

Table 5. Statistical descriptive of local weather data.

Table 5. Statistical descriptive of local weather data. 

Weather Parameter  Number of Data Point 
Weather Number of Data Min  Mean  Max 
Min Mean Max
Temperature  Parameter34,826  Point −9.02  11.23  37.31 
Windspeed  Temperature34,826  34,826 0.00  −9.02 2.59 
11.23 10.39 
37.31
Irradiation  Windspeed 34,826  34,826 0.00  0.00 2.59
135.33  10.39
1310.83 
Irradiation 34,826 0.00 135.33 1310.83

6. Result and Discussion 
6. Result and Discussion
6.1. A Comparison of the Hybrid Deep Learning Model for Predicting the Direction of Power 
6.1. A Comparison of the Hybrid Deep Learning Model for Predicting the Direction of Power Flow
Flow of Each Line Based on Real Power Measurement Data Only 
of Each Line Based on Real Power Measurement Data Only
In this subsection, we present the simulation results for predicting the direction of 
In this subsection, we present the simulation results for predicting the direction of
the power flow of each feeder in the studied grid cluster, based solely on real power meas‐
the power flow of each feeder in the studied grid cluster, based solely on real power
urement data (dataset group 1). For this simulation, the two proposed HDL models were 
measurement data (dataset group 1). For this simulation, the two proposed HDL models
used together with the baseline models. To compare the performance of all the deep learn‐
were used together with the baseline models. To compare the performance of all the deep
ing  models, 
learning we  considered 
models, the the
we considered duration  of of
duration the 
thetraining 
trainingperiod 
periodand 
and the  performance 
the performance
evaluation results based on various metrics, such as RMSE, MAE and R2 . During the
training period, all deep learning models were re-trained for different sub-datasets with the
  same fitting configurations. The reason for this is that the developed method is to predict
the direction of power flow of a single line in real time based on the other existing lines of
the network cluster. For example, if the DL model wants to predict the power flow in line 1,
it needs reference values for the power flow of lines 2, 3, 4, 5 and 6. This implementation
also applies to other lines if they are to be predicted. Therefore, six individual partial
prediction models of the individual deep learning models were created for each line.
As for the comparison of the training time of all the deep learning models, we can see
this in Table 6. The proposed HDL model of CNN-LSTM always has shorter training times
compared to the proposed model of LSTM-CNN in all the line-of-grid clusters studied,
although they have similar constructed layers and parameters used. Basically, CNNs are
designed to be faster because the computations in CNNs can be performed in parallel,
whereas LSTMs have to be processed sequentially because the next step depends on the
previous one. This can be illustrated in Figure 9 (right side), where the CNN trained with
the group 1 dataset (dataset containing only power measurements) has a faster training
time than other deep learning models. Therefore, the CNN network placed in the first layer
of the proposed HDL model can lead to various complexity reductions by focusing on the
most important features. The use of convolutional layers leads to a reduction in the size of
the tensor and the use of pooling layers also leads to a further reduction. This is one of the
reasons why the model CNN-LSTM can be faster than LSTM-CNN.
 Sensors 2023, 23, 901 13 of 20

Table 6. Training time of the deep learning models.

Training Time (Second)

Dataset DL Model
Line 1 Line 2 Line 3 Line 4 Line 5 Line 6
LSTM 168.849 169.043 167.98 158.55 158.56 222.20
CNN 49.522 48.855 46.316 47.85 46.11 49.22
Power data
CNN-LSTM 102.955 99.872 103.379 109.088 95.847 105.188
LSTM-CNN 182.87 188.738 192.276 185.572 186.291 190.605
LSTM 226.62 237.25 215.84 209.41 203.17 204.47
Power + CNN
Sensors 2023, 22, x FOR PEER REVIEW  57.76 52.95 52.60 51.76 49.07 50.24
15 of 22 
 
Weather data CNN-LSTM 129.591 146.837 121.648 134.741 131.83 119.886
LSTM-CNN 249.918 246.845 243.63 250.825 229.769 228.971

 
Figure 9. Training duration time of all the deep learning models.
Figure 9. Training duration time of all the deep learning models. 
With regard to the comparison of assessment performances. In Table 7, we see that
the two proposed HDL models are quite competitive in predicting power flow. Therefore,
we also need to compare the two HDL models with the baseline models. According to the
metrics RMSE, MAE, and R2 , the LSTM-CNN model performs better than the CNN-LSTM
model in predicting the power flow on lines 1, 2, 3, and 5, while the CNN-LSTM model
performs better only on lines 4 and 6. However, the overall comparison with the baseline
models, the metrics RMSE and R2 shows the LSTM model has the best performance among
all the deep learning models in predicting the power flow on lines 1, 3, 4, and 5, while
the proposed model LSTM-CNN has the best performance in predicting lines 2 and CNN-
LSTM for line 6. For the metric MAE, the LSTM model performs better than all other
models in predicting lines 3, 4, and 5, while the model LSTM-CNN performs better than all
other models in predicting lines 1 and 2 and the model CNN-LSTM performs better than
all other models in predicting line 6. Based on these simulation results, the HDL model
does not perform better than the single model of LSTM in predicting the power flow of the
network cluster under study.
In this prediction simulation, the direction of the power flow can be determined by the
value of the power flow itself. If the power value is below zero, the direction of the power
flow is from the network cluster to the busbar, while if it is above zero, it flows from the
Sensors 2023, 23, 901 14 of 20

busbar to the network cluster. The power flow prediction results of all the deep learning
models can be seen in Figure 10. This figure describes the prediction results for the test
Sensors 2023, 22, x FOR PEER REVIEW 
dataset covering the period from 7 November 2019 at 09:00 to 7 November 2019 at 20:45.15 of 22 
As
 
we can see, all deep learning models are generally equally good at following the original
value of the power flow measurement (purple line). In addition, We can also see in this
figure that most of the active power values (purple lines) on the feeder lines 3 and 4 are
always above 0. This indicates that the direction of power flow during this period tends to
be towards the grid cluster. On feeder lines 1 and 2, on the other hand, the power flow is
always below zero, indicating that there is a power surplus from the grid cluster and the
power flow is directly to the busbar. However, on closer inspection, the original power
flow pattern (purple line) on lines 1 and 2 are similar because these lines are connected in
parallel. Line 3 is also connected in parallel with line 4, as is line 5 to line 6.

Table 7. Performance evaluation of the deep learning models based on power flow data.

Metric Line Predicted

DL Model
Evaluation Line 1 Line 2 Line 3 Line 4 Line 5 Line 6
LSTM 4.694 8.195 2.698 3.47 4.33 2.81
RMSE CNN 4.823 7.299 2.931 4.16 5.51 2.94
(MW) CNN-LSTM 5.31 6.986 3.84 3.85 4.815 2.725
LSTM-CNN 4.922 6.917 2.991 4.118 4.747 2.877
LSTM 2.643 5.521 1.766 2.21 2.76 2.23
MAE CNN 2.412 3.318 1.811 2.81 3.97 2.36
(MW) CNN-LSTM 3.078 2.917 2.921 2.613 3.265 2.152
LSTM-CNN 2.408 2.809 1.838 2.745 3.175 2.281
LSTM 0.977 0.927 0.982 0.98 0.82 0.92
CNN 0.976 0.942 0.978 0.97 0.71 0.92
R2  
CNN-LSTM 0.97 0.947 0.963 0.975 0.778 0.93
LSTM-CNN Figure 9. Training duration time of all the deep learning models. 
0.975 0.948 0.978 0.971 0.785 0.92

 
Figure 10. Prediction results of the power flow direction based on real power measurement data 
Figure 10. Prediction results of the power flow direction based on real power measurement data only.
only. 

   
Sensors 2023, 23, 901 15 of 20

6.2. A Comparison of the Hybrid Deep Learning Model for Predicting the Direction of Power Flow
of Each Line Based on Real Power Measurement Data and Local Weather Data
In this sub-section, we show the simulation and test results of the proposed HDL
model for power flow prediction based on the input dataset containing power values along
with weather parameters (dataset group 2). The procedure used in the simulation to predict
the direction of power flow on the single feeder is exactly the same as in the previous
subsection. The only difference is the dataset used, as the purpose of this simulation was to
determine the extent to which weather data can affect the power flow prediction results.
Thus, if the model wants to predict the power flow in line 1, it needs reference values for
the power flow of lines 2, 3, 4, 5 and 6, as well as three additional weather parameters,
including air temperature, wind speed and solar irradiation
The competition between the CNN-LSTM and LSTM-CNN is also quite close in this
simulation, as Table 8 shows. Judging by the comparison of the training time (see Table 6),
the model CNN-LSTM always has a rather short training time compared to LSTM-CNN.
This is also evidenced by the training time in the previous section. Therefore, it can be
assumed that the training time of the CNN-LSTM model is indeed faster than that of the
LSTM-CNN model. However, the CNN model still has the fastest training time compared
to all models of DL (see Figure 9, left side). As for the comparison of the datasets used,
Figure 9 shows in detail that the training time of all models becomes longer when the
training datasets becomes larger. This can be seen when the input dataset for training
consists of power flow and local weather data. All models tend to have a longer training
time than the input dataset consisting only of power flow data. From this experiment, it
can be concluded that the size of the input dataset is a factor that significantly influences
the duration of the model training time.

Table 8. Performance evaluation of the deep learning models based on the power flow and weather data.

Metric Line Predicted

DL Model
Evaluation Line 1 Line 2 Line 3 Line 4 Line 5 Line 6
LSTM 5.897 7.475 3.075 5.026 2.335 2.589
RMSE CNN 5.858 7.169 2.992 4.042 2.51 3.124
(MW) CNN-LSTM 6.553 7.232 3.149 4.095 2.849 2.573
LSTM-CNN 5.649 7.056 3.242 4.746 3.026 2.98
LSTM 2.936 3.735 2.074 3.056 1.552 2.021
MAE CNN 2.974 3.689 2.036 2.782 1.673 2.526
(MW) CNN-LSTM 4.015 2.871 2.231 2.686 2.052 2.006
LSTM-CNN 2.733 2.868 1.985 3.23 2.468 2.416
LSTM 0.963 0.94 0.976 0.957 0.948 0.935
CNN 0.964 0.945 0.978 0.972 0.94 0.905
R2 CNN-LSTM 0.955 0.944 0.975 0.971 0.922 0.935
LSTM-CNN 0.966 0.946 0.974 0.961 0.913 0.913

From a performance comparison perspective in this simulation, the CNN-LSTM

and LSTM-CNN models tend to have the same ability to predict power flow. According
to the metrics RMSE, MAE and R2 , LSTM-CNN performs very well in predicting power
flow in lines 1 and 2, while CNN-LSTM surprisingly performs better in predicting lines
3, 4, 5 and 6. Since the proposed HDL model has similar performance, we considered the
LSTM only and the CNN only as the baseline models in this simulation and compared
them with the proposed HDL model. The experimental results (see Table 8) show the
HDL model performs much better than the single model. Based on the MAE metric,
LSTM-CNN performs better in predicting lines 1, 2, and 3 compared to the other models.
Meanwhile, CNN-LSTM is excellent in predicting lines 4 and 6 compared to the other
models, and the last one, the LSTM model, is the only good at predicting line 5. From
the RMSE and R2 perspective, the LSTM-CNN performs better compared to the other
models in predicting the power flow at lines 1 and 2, while CNN-LSTM is outstanding
Sensors 2023, 23, 901 16 of 20
Sensors 2023, 22, x FOR PEER REVIEW  17 of 22 
 

in predicting line 6 only. Unexpectedly, the CNN model performs well in predicting
one, the LSTM model, is the only good at predicting line 5. From the RMSE and R
lines 3 and 4 compared to the other models. Meanwhile in line 5, only LSTM has the2 per‐ best
spective, the LSTM‐CNN performs better compared to the other models in predicting the 
performance to predict power flow.
power flow at lines 1 and 2, while CNN‐LSTM is outstanding in predicting line 6 only. 
From the experiments conducted, the performance of the proposed HDL model and
Unexpectedly, the CNN model performs well in predicting lines 3 and 4 compared to the 
the baseline models tend to differ when two different datasets are used. In the previous
simulation, all models were trained with the dataset of group 1 (dataset containing only
other models. Meanwhile in line 5, only LSTM has the best performance to predict power 
power flow data). The evaluation results show that the HDL model is not superior to the
flow. 
baseline models. In contrast, the performance of the HDL model is good compared to
From the experiments conducted, the performance of the proposed HDL model and 
the baseline models in the simulation where all models were trained with the datasets of
the baseline models tend to differ when two different datasets are used. In the previous 
group 2 (containing power flow and weather parameters). Using these simulation results
simulation, all models were trained with the dataset of group 1 (dataset containing only 
(see Tables 7 and 8), we can check the performance of each model when trained with two
power flow data). The evaluation results show that the HDL model is not superior to the 
different datasets in terms of size and input parameters. The results of the performance
baseline models. In contrast, the performance of the HDL model is good compared to the 
comparison
baseline  between
models  thesimulation 
in  the  models trained with
where  all  the dataset
models  of group
were  1 with 
trained  (orangethe bar) and the
datasets  of 
models trained with the dataset of group 2 (blue bar) can be seen in Figures 11–13. From
group 2 (containing power flow and weather parameters). Using these simulation results 
these figures, it can be seen that all the evaluation metrics (RMSE, MAE, and R2 ) show that
(see Tables 7 and 8), we can check the performance of each model when trained with two 
the addition of weather parameters to the training dataset can affect the performance of all
different datasets in terms of size and input parameters. The results of the performance 
the models but does not have a significant impact. On closer inspection, the weather data
comparison between the models trained with the dataset of group 1 (orange bar) and the 
can improve the performance of the LSTM in predicting the power flow in lines 2, 5, and 6.
models trained with the dataset of group 2 (blue bar) can be seen in Figures 11–13. From 
While the CNN model has a better performance in lines 2, 4, and 5.
these figures, it can be seen that all the evaluation metrics (RMSE, MAE, and R 2) show that 
For the HDL model, adding weather parameters to the dataset used in the
the addition of weather parameters to the training dataset can affect the performance of  training
phase has a good impact on model performance. On closer inspection (see Figures
all the models but does not have a significant impact. On closer inspection, the weather  11–13),
the evaluation results show that the model LSTM-CNN with additional weather parameters
data can improve the performance of the LSTM in predicting the power flow in lines 2, 5, 
gives a better prediction of the power flow in lines 2, 3, 5 and 6. In contrast, the CNN-LSTM
and 6. While the CNN model has a better performance in lines 2, 4, and 5. 
model only has an effect on line 5.

 
Figure 11. Performance comparison based on the RMSE metric.
Figure 11. Performance comparison based on the RMSE metric. 

 
Sensors 2023, 22, x FOR PEER REVIEW  18 of 22 
 
Sensors 2023, 22, x FOR PEER REVIEW  18 of 22 
  Sensors 2023, 23, 901 17 of 20

 
Figure 12. Performance comparison based on the MAE metric.   
Figure 12. Performance comparison based on the MAE metric.
Figure 12. Performance comparison based on the MAE metric. 

 
Figure 13. Performance comparison based on the R2 metric.
Figure 13. Performance comparison based on the R2 metric.   
Figure 13. Performance comparison based on the R
The results of predicting the power flow is2 metric. 
very important in determining its direction.
For the HDL model, adding weather parameters to the dataset used in the training 
In this sub-section. The results of predicting the power flow of all the deep learning models
phase has a good impact on model performance. On closer inspection (see Figures 11–13), 
For the HDL model, adding weather parameters to the dataset used in the training 
trained with the dataset of group 2 (the dataset containing power values and weather
the evaluation results show that the model LSTM‐CNN with additional weather parame‐
phase has a good impact on model performance. On closer inspection (see Figures 11–13), 
parameters) is shown in Figure 14. The test dataset used in this simulation covers the
ters gives a better prediction of the power flow in lines 2, 3, 5 and 6. In contrast, the CNN‐
the evaluation results show that the model LSTM‐CNN with additional weather parame‐
period from 8 November 2019 at 09:00 to 8 November 2019 at 20:45. Looking closely at
LSTM model only has an effect on line 5. 
ters gives a better prediction of the power flow in lines 2, 3, 5 and 6. In contrast, the CNN‐
the prediction results, all the trained models are generally equally good at following the
LSTM model only has an effect on line 5. 
original value of the power flow measurement (purple line). As we can see from feeder
lines 3, 4, 5 and 6 (see Figure 14). The power flow is always above zero, which indicates

 
 
 tion. In this sub‐section. The results of predicting the power flow of all the deep learning 
models  trained  with  the  dataset  of  group  2  (the  dataset  containing  power  values  and 
weather parameters) is shown in Figure 14. The test dataset used in this simulation covers 
the period from 8 November 2019 at 09:00 to 8 November 2019 at 20:45. Looking closely 
at the prediction results, all the trained models are generally equally good at following 
Sensors 2023, 23, 901 18 of 20
the original value of the power flow measurement (purple line). As we can see from feeder 
lines 3, 4, 5 and 6 (see Figure 14). The power flow is always above zero, which indicates 
that the direction of the power flow is from the busbar to the cluster grid. This is because 
that the direction of the power flow is from the busbar to the cluster grid. This is because
there is a high load demand in the grid cluster. 
there is a high load demand in the grid cluster.

 
Figure
Figure  14. Prediction results of 
14. Prediction  results of the 
the power 
power flow 
flow direction
direction  based
based  on
on  real
real power
power  measurement
measurement  and
and 
weather data. 
weather data.

7. Conclusions
7. Conclusions 
As the growth and deployment of renewable energy systems increases rapidly, the
As the growth and deployment of renewable energy systems increases rapidly, the 
energy exchange becomes more complex. This is because consumers act as prosumers,
energy  exchange  becomes  more  complex.  This  is  because  consumers  act  as  prosumers, 
meaning they have the ability to generate electricity and synchronize with the grid. They
meaning they have the ability to generate electricity and synchronize with the grid. They 
also tend to have different types of loads that are predominantly inductive or capacitive.
also tend to have different types of loads that are predominantly inductive or capacitive. 
This phenomenon will draw active energy from the system or towards the system. There-
This phenomenon will draw active energy from the system or towards the system. There‐
fore, it is necessary to use the correct method to determine exactly how much active and
fore, it is necessary to use the correct method to determine exactly how much active and 
reactive energy is generated in the power system under export or import conditions. Know-
reactive  energy  is  generated  in  the  power  system  under  export  or  import  conditions. 
ing the direction of power flow during energy exchange is very important for distribution
Knowing the direction of power flow during energy exchange is very important for dis‐
companies as it can help control the energy generated or distributed during the energy
tribution companies as it can help control the energy generated or distributed during the 
exchange between different but interconnected utilities.
energy exchange between different but interconnected utilities. 
In this study, two of the most popular hybrid deep learning models were proposed
In this study, two of the most popular hybrid deep learning models were proposed 
and compared with the baseline models (CNN only and LSTM only) to predict the direction
and compared with the baseline models (CNN only and LSTM only) to predict the direc‐
of power flow on each line of a network cluster. To compare the performance of all the
tion of power flow on each line of a network cluster. To compare the performance of all 
deep learning models, we have considered the duration of the training period and the
the deep learning models, we have considered the duration of the training period and the 
performance evaluation results based on various metrics, such as RMSE, MAE and R2 . In
performance evaluation results based on various metrics, such as RMSE, MAE and R 2. In 
the training process, two types of datasets were used to test the capability of the proposed
HDL model. The first type was a dataset containing real power measurement data, and the
second type was a group of datasets containing real power measurement and local weather
data. The purpose of dividing this group of datasets was to see if the size of the dataset
 
and the weather data affected the performance of the proposed model.
The experimental results show that the weather parameters in the dataset can increase
the size of the datasets and increase the training time of all the models. Therefore, the size
of the input dataset may affect the training time of the model. In terms of performance
evaluation, the proposed HDL model did not perform better than the baseline models in
predicting the power flow of the studied grid cluster when trained with the power flow
Sensors 2023, 23, 901 19 of 20

data only. However, in contrast, the proposed HDL models showed good performance
compared to the baseline models when trained with a dataset that included additional
weather parameters. In this study, the metric evaluation of RMSE, MAE and R2 shows
that both proposed HDL models can reach a small error under certain conditions, so it is
still relatively challenging to determine the best HDL model for predicting the direction of
power flow in the investigated network cluster.

Author Contributions: Conceptualization, F.A. and Y.L.; methodology, F.A. and Y.L.; software, F.A.,
Y.L. and V.S.; validation, F.A., Y.L. and V.S.; validation, F.A.; formal analysis, V.S.; investigation, Y.L.;
resources, P.J.; data curation, Y.L. and P.J.; writing—original draft preparation, F.A.; writing—review
and editing, F.A. and V.S.; visualization, F.A.; supervision, V.S. and P.J.; project administration, V.S.
All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data are not publicly available due to the policy of the associate
company.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wang, Q.; Dong, Z.; Li, R.; Wang, L. Renewable energy and economic growth: New insight from country risks. Energy 2022, 238,
122018. [CrossRef]
2. Aslam, M.; Lee, J.-M.; Kim, H.-S.; Lee, S.-J.; Hong, S. Deep Learning Models for Long-Term Solar Radiation Forecasting
Considering Microgrid Installation: A Comparative Study. Energies 2019, 13, 147. [CrossRef]
3. Brauns, K.; Scholz, C.; Schultz, A.; Baier, A.; Jost, D. Vertical power flow forecast with LSTMs using regular training update
strategies. Energy AI 2021, 8, 100143. [CrossRef]
4. Li, Y.; Janik, P.; Schwarz, H.; Pfeiffer, K. Proposal of a regional grid cluster model for analysis of electrical power net-work
performance. Arch. Electr. Eng. 2022, 71, 601–613.
5. Suresh, G.; Prasad, D.; Gopila, M. An efficient approach based power flow management in smart grid system with hybrid
renewable energy sources. Renew. Energy Focus 2021, 39, 110–122. [CrossRef]
6. Aslam, S.; Herodotou, H.; Mohsin, S.M.; Javaid, N.; Ashraf, N.; Aslam, S. A survey on deep learning methods for power load and
renewable energy forecasting in smart microgrids. Renew. Sustain. Energy Rev. 2021, 144, 110992. [CrossRef]
7. Zhou, H.; Wang, S.; Miao, Z.; He, C.; Liu, S. Review of The Application of Deep Learning in Fault Diagnosis. Chin. Control Conf.
CCC 2019, 2019, 4951–4955. [CrossRef]
8. Aksan, F.; Janik, P.; Suresh, V.; Leonowicz, Z. Review of the application of deep learning for fault detection in wind turbine. In
Proceedings of the 2022 IEEE International Conference on Environment and Electrical Engineering and 2022 IEEE Industrial and
Commercial Power Systems Europe (EEEIC/I&CPS Europe), Prague, Czech Republic, 28 June 2022—1 July 2022; pp. 1–6. [CrossRef]
9. Alkhayat, G.; Mehmood, R. A review and taxonomy of wind and solar energy forecasting methods based on deep learning.
Energy AI 2021, 4, 100060. [CrossRef]
10. Cecaj, A.; Lippi, M.; Mamei, M.; Zambonelli, F. Comparing Deep Learning and Statistical Methods in Forecasting Crowd
Distribution from Aggregated Mobile Phone Data. Appl. Sci. 2020, 10, 6580. [CrossRef]
11. Fallah, S.N.; Deo, R.C.; Shojafar, M.; Conti, M.; Shamshirband, S. Computational Intelligence Approaches for Energy Load Forecasting
in Smart Energy Management Grids: State of the Art, Future Challenges, and Research Directions. Energies 2018, 11, 596. [CrossRef]
12. Wu, Y.-X.; Wu, Q.-B.; Zhu, J.-Q. Data-driven wind speed forecasting using deep feature extraction and LSTM. IET Renew. Power
Gener. 2019, 13, 2062–2069. [CrossRef]
13. Yu, C.; Li, Y.; Bao, Y.; Tang, H.; Zhai, G. A novel framework for wind speed prediction based on recurrent neural networks and
support vector machine. Energy Convers. Manag. 2018, 178, 137–145. [CrossRef]
14. Tong, X.; Wang, J.; Zhang, C.; Wu, T.; Wang, H.; Wang, Y. LS-LSTM-AE: Power load forecasting via Long-Short series features and
LSTM-Autoencoder. Energy Rep. 2022, 8, 596–603. [CrossRef]
15. Wang, F.; Xuan, Z.; Zhen, Z.; Li, K.; Wang, T.; Shi, M. A day-ahead PV power forecasting method based on LSTM-RNN model and
time correlation modification under partial daily pattern prediction framework. Energy Convers. Manag. 2020, 212, 112766. [CrossRef]
16. Suresh, V.; Aksan, F.; Janik, P.; Sikorski, T.; Revathi, B.S. Probabilistic LSTM-Autoencoder Based Hour-Ahead Solar Power Forecasting
Model for Intra-Day Electricity Market Participation: A Polish Case Study. IEEE Access 2022, 10, 110628–110638. [CrossRef]
17. Kumar, S.; Hussain, L.; Banarjee, S.; Reza, M. Energy Load Forecasting using Deep Learning Approach-LSTM and GRU in Spark
Cluster. In Proceedings of the 2018 Fifth International Conference on Emerging Applications of Information Technology (EAIT),
Kolkata, India, 12–13 January 2018; pp. 1–4. [CrossRef]
Sensors 2023, 23, 901 20 of 20

18. Feng, C.; Zhang, J. SolarNet: A sky image-based deep convolutional neural network for intra-hour solar forecasting. Sol. Energy
2020, 204, 71–78. [CrossRef]
19. Suresh, V.; Janik, P.; Rezmer, J.; Leonowicz, Z. Forecasting Solar PV Output Using Convolutional Neural Networks with a Sliding
Window Algorithm. Energies 2020, 13, 723. [CrossRef]
20. Fu, J.; Chu, J.; Guo, P.; Chen, Z. Condition Monitoring of Wind Turbine Gearbox Bearing Based on Deep Learning Model. IEEE
Access 2019, 7, 57078–57087. [CrossRef]
21. Zang, H.; Liu, L.; Sun, L.; Cheng, L.; Wei, Z.; Sun, G. Short-term global horizontal irradiance forecasting based on a hybrid
CNN-LSTM model with spatiotemporal correlations. Renew. Energy 2020, 160, 26–41. [CrossRef]
22. Park, J.; Kodaira, D.; Agyeman, K.; Jyung, T.; Han, S. Adaptive Power Flow Prediction Based on Machine Learning. Energies 2021,
14, 3842. [CrossRef]
23. Schäfer, F.; Menke, J.-H.; Braun, M. Prediction of power flow results in time-series-based planning with artificial neural networks
and data pre-processing. CIRED—Open Access Proc. J. 2020, 2020, 74–77. [CrossRef]
24. Song, J.; Zhang, L.; Xue, G.; Ma, Y.; Gao, S.; Jiang, Q. Predicting hourly heating load in a district heating system based on a hybrid
CNN-LSTM model. Energy Build. 2021, 243, 110998. [CrossRef]
25. Agga, A.; Abbou, A.; Labbadi, M.; El Houm, Y.; Ali, I.H.O. CNN-LSTM: An efficient hybrid deep learning architecture for
predicting short-term photovoltaic power production. Electr. Power Syst. Res. 2022, 208, 107908. [CrossRef]
26. Farsi, B.; Amayri, M.; Bouguila, N.; Eicker, U. On Short-Term Load Forecasting Using Machine Learning Techniques and a Novel
Parallel Deep LSTM-CNN Approach. IEEE Access 2021, 9, 31191–31212. [CrossRef]
27. Li, C.; Hu, R.; Hsu, C.-Y.; Han, Y. Short-term Power Load Forecasting based on Feature Fusion of Parallel LSTM-CNN. In
Proceedings of the 2022 IEEE 4th International Conference on Power, Intelligent Computing and Systems (ICPICS), Shenyang,
China, 29–31 July 2022; pp. 448–452. [CrossRef]
28. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef] [PubMed]
29. Tulensalo, J.; Seppänen, J.; Ilin, A. An LSTM model for power grid loss prediction. Electr. Power Syst. Res. 2020, 189, 106823.
[CrossRef]
30. Qing, X.; Niu, Y. Hourly day-ahead solar irradiance prediction using weather forecasts by LSTM. Energy 2018, 148, 461–468.
[CrossRef]
31. Ghosh, A.; Sufian, A.; Sultana, F.; Chakrabarti, A.; De, D. Fundamental Concepts of Convolutional Neural Network. Intell. Syst.
Ref. Libr. 2019, 172, 519–567. [CrossRef]
32. Habeck, C.; Gazes, Y.; Razlighi, Q.; Stern, Y. Cortical thickness and its associations with age, total cognition and education across
the adult lifespan. PLoS ONE 2020, 15, e0230298. [CrossRef]
33. Mishra, B.; Shahi, T.B. Deep learning-based framework for spatiotemporal data fusion: An instance of Landsat 8 and Sentinel 2
NDVI. J. Appl. Remote. Sens. 2021, 15, 034520. [CrossRef]
34. Plakias, S.; Boutalis, Y.S. Fault detection and identification of rolling element bearings with Attentive Dense CNN. Neurocomputing
2020, 405, 208–217. [CrossRef]
35. Chen, L.; Xu, G.; Zhang, Q.; Zhang, X. Learning deep representation of imbalanced SCADA data for fault detection of wind
turbines. Measurement 2019, 139, 370–379. [CrossRef]
36. Aksan, F.; Jasiński, M.; Sikorski, T.; Kaczorowska, D.; Rezmer, J.; Suresh, V.; Leonowicz, Z.; Kostyła, P.; Szymańda, J.; Janik, P.
Clustering Methods for Power Quality Measurements in Virtual Power Plant. Energies 2021, 14, 5902. [CrossRef]
37. Lee, T.; Singh, V.P.; Cho, K.H. Deep Learning for Time Series. Water Sci. Technol. Libr. 2021, 99, 107–131. [CrossRef]
38. Wang, K.; Qi, X.; Liu, H. A comparison of day-ahead photovoltaic power forecasting models based on deep learning neural
network. Appl. Energy 2019, 251, 113315. [CrossRef]
39. Wang, K.; Qi, X.; Liu, H. Photovoltaic power forecasting based LSTM-Convolutional Network. Energy 2019, 189, 116225. [CrossRef]
40. Memarzadeh, G.; Keynia, F. A new short-term wind speed forecasting method based on fine-tuned LSTM neural network and
optimal input sets. Energy Convers. Manag. 2020, 213, 112824. [CrossRef]
41. Rahimilarki, R.; Gao, Z.; Jin, N.; Zhang, A. Time-series Deep Learning Fault Detection with the Application of Wind Turbine
Benchmark. IEEE Int. Conf. Ind. Inform. 2019, 1, 1337–1342. [CrossRef]
42. Lu, X.; Lin, P.; Cheng, S.; Lin, Y.; Chen, Z.; Wu, L.; Zheng, Q. Fault diagnosis for photovoltaic array based on convolutional neural
network and electrical time series graph. Energy Convers. Manag. 2019, 196, 950–965. [CrossRef]
43. Hwang, H.P.-C.; Ku, C.C.-Y.; Chan, J.C.-C. Detection of Malfunctioning Photovoltaic Modules Based on Machine Learning
Algorithms. IEEE Access 2021, 9, 37210–37219. [CrossRef]
44. Jahangir, H.; Golkar, M.A.; Alhameli, F.; Mazouz, A.; Ahmadian, A.; Elkamel, A. Short-term wind speed forecasting framework
based on stacked denoising auto-encoders with rough ANN. Sustain. Energy Technol. Assessments 2020, 38, 100601. [CrossRef]
45. Hong, Y.-Y.; Rioflorido, C.L.P.P. A hybrid deep learning-based neural network for 24-h ahead wind power forecasting. Appl.
Energy 2019, 250, 530–539. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.