Renewable and Sustainable Energy Reviews 172 (2023) 113046

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

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

A hybrid framework for forecasting power generation of multiple

renewable energy sources
Jianqin Zheng a, 1, Jian Du a, 1, Bohong Wang b, *, Jiří Jaromír Klemeš c, Qi Liao a,
Yongtu Liang a, **
a
National Engineering Laboratory for Pipeline Safety/ MOE Key Laboratory of Petroleum Engineering/ Beijing Key Laboratory of Urban Oil and Gas Distribution
Technology, China University of Petroleum-Beijing, Fuxue Road No. 18, Changping District, Beijing, 102249, PR China
b
National & Local Joint Engineering Research Center of Harbor Oil & Gas Storage and Transportation Technology/Zhejiang Key Laboratory of Petrochemical
Environmental Pollution Control, Zhejiang Ocean University, No.1 Haida South Road, Zhoushan, 316022, PR China
c
Sustainable Process Integration Laboratory – SPIL, NETME Centre, Faculty of Mechanical Engineering, Brno University of Technology - VUT BRNO, Technická 2896/2,
616 69, Brno, Czech Republic

A R T I C L E I N F O A B S T R A C T

Keywords: The accurate power generation forecast of multiple renewable energy sources is significant for the power
Renewable energy system scheduling of renewable energy systems. However, previous studies focused more on the prediction of a single
Power generation energy source, ignoring the relationship among different energy sources, and failing to predict accurate power
Forecasting
generation for all energy sources simultaneously. This paper proposes a hybrid framework for the power gen­
Hybrid framework
Long short-term memory
eration forecast of multiple renewable energy sources to overcome deficiencies. A Convolutional Neural Network
(CNN) is developed to extract the local correlations among multiple energy sources, the Attention-based Long
Short-Term Memory (A-LSTM) network is developed to capture the nonlinear time-series characteristics of
weather conditions and individual energy, and the Auto-Regression model is applied to extract the linear time-
series characteristics of each energy source. The accuracy and practicality of the proposed method are verified by
taking a renewable energy system as an example. The results show that the hybrid framework is more accurate
than other advanced models, such as artificial neural networks and decision trees. Mean absolute errors of the
proposed method are reduced by 13.4%, 22.9%, and 27.1% for solar PV, solar thermal, and wind power
compared with A-LSTM. The sensitivity analysis has been conducted to test the effectiveness of each component
of the proposed hybrid framework to prove the significance of energy correlation patterns with higher accuracy
and stability compared with the other two patterns.

with the promotion of renewable energy technology and the imple­

mentation of a near “zero carbon emission” policy, green development is
1. Introduction the foundation of people’s livelihood [5]. Against this background,
many countries paid attention to renewable energy systems (RESs) to
1.1. Background utilise renewable energy sources for electricity generation [6]. Nearly
60% of all new power generation capacities in 2040 would come from
Along with the rapid development of the social economy, the energy renewables, stated to International Energy Agency (IEA) [7].
demand has increased significantly to fulfil daily human demands and Due to the proposal of the smart microgrid, the application of RESs
activities [1]. The growing dependence on energy has led to a significant has increased greatly, considering that the high penetration of RESs has
increase in power generation. Traditional fossil fuels have caused a lot of advantages, for example, low environmental pollution, energy
serious environmental pollution [2], such as global warming and air supply security, and sustainable development realisation [8]. RESs
pollution [3]. Greenhouse gas emissions generated by these fossil fuels generally consist of energy generation, energy storage components, and
contribute to global warming and climate change [4]. At the same time,

* Corresponding author.
** Corresponding author.
E-mail addresses: [email protected] (B. Wang), [email protected] (Y. Liang).
1
Both authors contributed equally to this work and should be considered co-first authors.

https://doi.org/10.1016/j.rser.2022.113046
Received 7 February 2022; Received in revised form 11 September 2022; Accepted 7 November 2022
Available online 16 November 2022
1364-0321/© 2022 Elsevier Ltd. All rights reserved.
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Nomenclature ht− 1 The output of state t − 1

ht The output of state t
Abbreviations it The output of the input gate
A–CNN–LSTM Attention-based convolution neural network long- klij The weights in l th layer between i th input feature map and
short term memory j th feature map
ANN Artificial neural network Mj The selection of input feature maps
AR Auto-regressive N The total number of data
A-LSTM Attention-based LSTM u, w The weighting coefficients of the attention layer
ARMA Auto-regressive moving average wi The coefficient of regressors
CI Confidence index wn The weather condition variables
CNN Convolutional neural network (Wf , bf ) The weights and biases of the forget gates
ELM Extreme learning machine (Wi , bi ) The weights and biases of the input gates
IEA International Energy Agency (Wo , bo ) The weights and biases of the output gates
KF Kalman filter
xi
̃ The normalised energy data
LSTM Long short-term memory
xi The input original energy data
MAE Mean absolute error
xmin The minimum value of original energy data
MSE Mean squared error
xmax The maximum value of original energy data
MLP Multilayer perceptron
xcnn The input multiple energy data of the CNN network
NWP Numerical weather prediction
xt The input data in state t
RES Renewable energy system
xl−i 1 The feature map of the (l − 1) layer
RNN Recurrent neural network
RMSE Root mean squared error xlj The output feature map
SVM Support vector machine xlstm The input matrix data of the LSTM network
XGB Extreme gradient boost yi The renewable energy resources
yi
̂ The predicted results
Variables yi The average values of original values
αi The weight of each hidden state εt The white noise which obeys the normal distribution
blj The bias of convolution layer
Parameters
b0 The intercept of AR model
down( ⋅) The max-pooling subsampling function
C
̃t The output of Tanh layer
f( ⋅) The activation function
Ct The cell state in state t
* The convolution operation
C The weighted feature
σ ( ⋅) The sigmoid activation function
ei The probability distribution of the attention layer
ft The output of the forget gate

loads, which can be used to supply some or all of the electricity needs for lower costs for balancing energy [19]. Consequently, this work studies
certain regions [9]. In RESs, the prediction of electricity generation from the power generation prediction in a RES considering the local corre­
renewable energy sources is important for both system operators and lation of multiple renewable energy resources, thus reducing uncer­
individual generators. For example, the Australian National Electricity tainty in the electricity supply.
Market requires power plants to submit dispatch offers up to 40 h ahead
and allow them to update the offers until 5 min before the dispatch is 1.2. Related work
made [10]. Reliable electricity generation prediction of RESs not only
improves the operational coordination of power systems to ensure en­ In recent years, many studies have been conducted to predict
ergy supply security, but it can also promote power dispatch efficiency renewable energy generation. However, most of them focused on a
for energy companies, optimise the power resource scheduling, and in­ single energy source, such as solar energy or wind energy. As one of the
crease economic benefits [11]. greatest potential sources, solar energy has attracted many countries to
In RESs, renewable energy sources can include biogas, biomass [12], develop large-scale solar power plants. Owing to expensive solar irra­
geothermal, small hydro, solar PV, solar thermal [13], and wind [14]. diance meters, some studies predicted solar irradiance rather than
The coordination of these sources of energy should be studied to in­ power generation. Wang et al. [20] proposed a method of solar irradi­
crease the accuracy of the multi-energy generation prediction [15]. The ance forecasting to achieve higher accuracy in minutely solar power
uncertainty exists in energy generation prediction, especially for solar forecasting by analysing the interaction relationship between sky image
and wind, as their generations are greatly influenced by the weather, and solar irradiance. Similarly, Alani et al. [21] proposed a hybrid
posing a threat to electricity scheduling for the smart microgrid. Either prediction model composed of a convolutional neural network (CNN)
PV modules or concentrated solar thermal can be applied to generate and multilayer perceptron (MLP) model to achieve 15 min-ahead fore­
electricity from solar irradiance [16] and wind power by wind farms. PV casting of the global irradiance using sky images. Hoyos-Gómez et al.
power is generated from global horizontal irradiance, while solar ther­ [22] introduced a pipeline for the one-day ahead forecasting of solar
mal is generated only from direct normal irradiance [17]. More and irradiance and insolation based on the historical data imputation stage
more wind farms have been installed, especially in China, America, and and four data-driven prediction stages. Huang et al. [23] developed a
Europe, with a number of wind sources [18]. However, both solar and novel multivariate hybrid deep neural model for 1-h-ahead solar irra­
wind power are intermittent owing to weather conditions, which lead to diance forecasting by considering climate influence, and the model was
the spatial-temporal correlation relationship. Accurate prediction of verified by actual data from four countries. Kumari and Toshniwal [24]
solar and wind energy is the basis to predict electricity supply by RESs to reviewed the deep learning-based solar irradiance forecasting models
reduce the supply risk of brownouts or even blackouts in the system and extensively and comprehensively to reduce the limitations of existing

2
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

machine learning models and improve prediction accuracy. hybrid forecasting frameworks have been developed. These hybrid
With the great development of measuring meters storing tremendous models combined the advantages of different methods and achieved
data and big data technology, statistical models and machine learning better performance than single models. Especially, decomposition and
models are used for PV power forecasting, such as auto-regressive feature selection focus on the processing of wind energy series [48], and
moving average (ARMA) [25], support vector machine (SVM) [26], ensemble learning can combine several predictors to form a more robust
and time series ensembles [27]. Deep learning-based methods play an model [49].
unprecedented role in improving the prediction accuracy of PV power to Through the review of solar PV, solar thermal, and wind power
tackle the deficiencies of traditional artificial intelligence modelling forecasting, the most popular approaches include physical, time series,
methods [28]. Li et al. [29] developed a hybrid deep learning model machine learning, and hybrid methods. Both solar energy forecasting
based on wavelet packet decomposition used to decompose the original and wind energy forecasting have been studied individually. However,
PV power series into sub-series and four independent long short-term solar and wind power generation exhibit spatial-temporal correlation
memory (LSTM) networks. Wang et al. [30] proposed a novel model because of their dependence on meteorological conditions. For example,
for day-ahead PV power forecasting integrating deep learning modelling solar energy changes with sunlight and irradiance, and solar power
and time-dependent principles under a partial daily pattern prediction generation increases with irradiance [50], and the main meteorological
framework which is used to guide the modification parameters. Ahmed driver of wind energy is wind speed, and wind power generation in­
et al. [31] reviewed and evaluated contemporary forecasting techniques creases with wind speed [51]. These studies mainly focused on a single
from the forecast horizon, timestamp, input correlation analysis, data energy source, ignoring the relationship among solar PV, solar thermal,
pre and post-processing, network optimisation, weather classification, and wind power and failing to improve the prediction accuracy of these
uncertainty quantification, and performance evaluations. In addition to three energy sources simultaneously. This work proposes a compre­
a certain region, a few studies considered the solar generation of mul­ hensive forecasting framework for solar PV, solar thermal, and wind
tiple regions. Zheng et al. [32] proposed an improved LSTM network power considering their indirect correlation based on data analysis and
optimised by particle swarm optimisation to predict the multi-region deep learning to fill the gap. As shown in Fig. 1, the main framework of
solar power output and illustrated the effectiveness of the method by this work can be split into three parts. The first part is input information,
a real area of Asia. Zhao et al. [33] considered spatial and meteorolog­ including correlation features among solar PV, solar thermal, and wind
ical characteristics of multi-region solar generation, then applied auto­ power, nonlinear time-series features of weather conditions and power
matic machine learning to generate the most suitable ensemble generation, and linear time-series features of individual renewable en­
prediction model with optimal parameters and genetic algorithm to ergy sources. The second part is the proposed hybrid forecasting method
process the candidate features. based on the input information, consisting of CNN, attention-based
Former studies mainly focused on the forecast of solar irradiance or LSTM (A-LSTM), and auto-regression (AR) to extract energy correla­
PV power generation, but only a few studies considered the prediction of tion patterns, nonlinear temporal patterns, and linear temporal patterns.
solar thermal, which is also an important part of solar energy. Wang The third part is comparison and analysis, including data preparation,
et al. [34] presented a hybrid method for forecasting solar thermal based power generation prediction, model comparison, and sensitivity anal­
on mechanism modelling and deep learning to identify the major ysis. For a RES, data is collected and divided into a training set, vali­
meteorological factors and couple the spatial-temporal characteristics dation set, and testing set. The proposed hybrid framework is trained
between meteorological factors. Rodat et al. [35] coupled a solar power based on the training set, and its hyperparameters are optimised by the
plant performance model with a meteorological model to predict the validation set. After that, the testing set is used for forecasting power
solar thermal output 24 h ahead in the case of a solar Fresnel power generation while comparing with other advanced prediction models
plant. Law et al. [36] presented a review of the direct normal irradiance trained only by information 2. Finally, sensitivity analysis is conducted
prediction accuracy of numerical weather prediction models, cloud to discuss the effectiveness of the hybrid framework without informa­
motion vectors, time series analysis methods, and hybrid methods to tion 1, or 2, or 3.
operate concentrated solar thermal plants. Although the current usage is
relatively low, concentrated solar thermal has an advantage in energy 1.3. Contributions of this work
storage because thermal energy storage is cheaper and more efficient
than electrical and mechanical energy storage options available to PV The contributions of this work can be concluded as follows:
[37].
Many studies proposed reliable wind speed and wind power fore­ (i) The potential correlations among solar PV, solar thermal, and
casting methods to improve the operational coordination of wind and wind energy are mined when forecasting power generation of
power systems [38]. Wind energy forecasting models have been them in RESs.
improved extensively, and they can be divided into deterministic pre­ (ii) A hybrid framework is proposed to forecast multiple energy
diction and uncertainty analysis [39]. The deterministic prediction generation, consisting of an A-LSTM layer capturing the
models mainly include physical, statistical, intelligent, and hybrid. As nonlinear temporal characteristics of weather conditions and
the representative of physical methods, numerical weather prediction power generation, a CNN layer mining the correlation of multiple
(NWP) applies the computer to solve the equations of fluid mechanics energy sources, and a linear layer considering the linear temporal
and thermodynamics [40]. The physical modes perform well in characteristics of each energy.
long-term forecasting but unsatisfactorily in short-term forecasting due (iii) A real RES is taken as an example to validate the performance of
to a large amount of calculation [41]. The statistical models are based on the proposed hybrid framework.
mathematical theory knowledge, including Kalman filter (KF) [42], (iv) Sensitivity analysis of different components combination is con­
Bayesian multiple kernel regression [43], copula theory [44], etc. The ducted to discuss the significance of each different information
statistical methods are widely used for simple time series forecasting, pattern.
but their ability cannot handle strongly nonlinear data [45]. Compared
with physical and statistical methods, intelligent forecasting models 1.4. The organisation of this work
have better accuracy due to their excellent capacity for analysing
nonstationary wind energy series. At present, intelligent methods The rest of this paper is organised as follows. Data preparation, the
perform irreplaceably in wind speed and wind power prediction, basic model components, the architecture of the hybrid prediction
including artificial neural networks (ANN) [46] and extreme learning framework, and evaluation metrics are introduced in section 2. In sec­
machines (ELM) [47]. In addition to the three single models, many tion 3, taking a real RES as a case study, the data description, model

3
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Fig. 1. The main framework of this work.

setting, prediction results, and sensitivity analysis are demonstrated and of [0, 1], using the Min-Max normalisation method as shown in Eq (1)
discussed in detail. Section 4 provides the conclusions and future work. [53]:
xi − xmin
2. Methodology xi =
̃ (1)
xmax − xmin

In this section, the hourly power generation prediction method of where ̃xi is the normalised energy data and xi denotes the input original
multiple renewable energy is established. Different deep learning energy data. xmin and xmax represent the minimum value and maximum
frameworks are utilised to extract different information patterns, and value of original energy data.
then a hybrid prediction framework with higher accuracy and efficiency After the input is prepared, the two-dimensional matrix is con­
is proposed to take different data features into account. structed by applying the approach of sliding window sampling.
Considering that there exists a local correlation pattern between
2.1. Data preparation different renewable energy resources, each column of the 2D matrix
contains not only the weather conditions but also the multiple renew­
It is significant to carry out data preparation for model development. able energy data. Besides, each row of the constructed matrix represents
Data preparation consists of several processes, such as data collection, different time steps.
data division, and data normalisation. In this study, the generation
power of different renewable energy can be collected from the RES, and 2.2. Model components
the corresponding weather conditions can be easily acquired from the
weather bureau. It is necessary to divide the collected data into three In this subsection, the basic principles of each component in the
different parts, i.e., training set, validation set, and testing set to ensure proposed framework are briefly described to develop the proposed
the robustness of the proposed hybrid framework [52]. The training set hybrid prediction framework.
is employed to update model parameters in the hybrid framework. The
validation set is formed to test the performance of the current model 2.2.1. Convolution neural network
after each epoch is finished. When the training process is completed, the As a special feedforward neural network [54], a typical CNN network
accuracy of the hybrid model will be tested by applying the testing set. consists of a convolution layer and a pooling layer. In general, the CNN
Data normalisation is necessary to be conducted to optimise the network attempt to extract hidden features of input data through a
model performance to accelerate the convergence rate of the hybrid layer-by-layer structure made up of a convolution layer and pooling
framework. The input data are normalised to the values within the range layer. This operating mode provides the CNN network with powerful

4
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

abilities of feature extraction and feature compression [55]., The CNN convolution layer and pooling layer are employed to extract the energy
network essentially realises a feature mapping without any explicit correlation pattern. Considering that the power generation possesses not
formal representation by achieving the representation of complex fea­ only nonlinear and linear time-series characteristics, it is significantly
tures from input to output. Each convolution layer in the CNN network important to extract comprehensive temporal information. As a conse­
employs various filters to obtain feature maps and generates the final quence, the memory block and attention block of A-LSTM are utilised to
output by applying feature maps eventually [56]. In this work, the extract the nonlinear temporal pattern of power generation and weather
convolution layer and pooling layer with the one-dimensional operation conditions. Besides, the linear time-series characteristics of a predicted
are utilised to extract energy correlation patterns between multiple variable are analysed through the AR method. Then the hybrid predic­
energy resources. tion framework is established for power generation prediction. In the
following content, each section of the proposed hybrid framework will
2.2.2. Attention-based long short-term memory be introduced in detail.
As an enhanced recurrent neural network (RNN), the LSTM network
possesses the capacity to store and remove temporal information auto­ 2.3.1. Energy correlation pattern extraction
matically. This ability contributes to overcoming the problem of The power generation prediction of multiple energy resources differs
gradient disappearance and gradient explosion of the RNN due to from the prediction of traditional single energy resources. Most of the
increasingly complex time-series features [57]. LSTM network contains previous studies aimed to approximate the relationship between inde­
a special form of repeated modules. Each module consists of three pendent variables and predicted variables. Nevertheless, there is an
control gates, named the forget gate, the input gate, and the output gate. interaction relation between multiple energy resources. Aiming to
When calculating the current output of the LSTM network, the compu­ extract the energy correlation pattern, the input multiple energy data
tation state of the previous LSTM cell and the current input state are are constructed as a two-dimensional matrix. Each column of the matrix
taken into consideration at the same time [58]. denotes a different energy resource, and each row represents a different
However, for the power generation prediction, each state of the time state. As shown in Fig. 3, given the input multiple energy data xcnn
previous time will have different degrees of influence on the subsequent (Eq. (2)), the convolution layer will create various feature maps
time, and it is important for model development to make the LSTM applying convolution filters. These feature maps contain complex local
network focus on useful information. Then the temporal attention correlations between multiple energy resources.
mechanism is integrated into the conventional LSTM network, called A-
xcnn = (y1 , y2 , y3 )t t ∈ [T − a, T − 1] (2)
LSTM [59].
where yi denotes different renewable energy resources. The process of
2.2.3. Autoregressive model
convolution layer can be formulated as follows:
AR model is a process of employing itself as the regression variable. ( )
When solving the time-series prediction problem, the variables in ∑
several previous periods are applied as the input of the linear regression xlj = f kijl ∗ xil− 1 + blj (3)
model [60]. Essentially speaking, the AR model is designed to approx­
i∈Mj

imate the linear relationship of the predicted variable between the

where klij denotes the weights in l th layer between i th input feature map
period {1, 2, ..., t − 1} and the current time t [61].
and j th feature map and blj is the bias. Mj represents the selection of
input feature maps. The feature map of the (l − 1) layer xl−i 1 conducts
2.3. Hybrid prediction framework
convolution operation × with klij . Subsequently, the output feature map
The structure of the proposed hybrid prediction framework is xlj is calculated through the activation function f( ⋅).
depicted in Fig. 2. To overcome the neglect of the local correlation The max-pooling layer is designed to reduce the dimensions of the
pattern that exists between different renewable energy resources, the

Fig. 2. The architecture of the proposed hybrid prediction model.

5
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Fig. 3. The structural diagram of CNN.

feature maps, aiming to solve the overfitting problem and shorten tiple energy resources to be predicted. As shown in Fig. 4, given the
training time. In the max-pooling layer, the number of CNN parameters input data xt in the state t, the forget gate first calculates ft , aiming to
is reduced. The operation process of the max-pooling layer can be rep­ decide what information ought to be retained according to the output
resented as follows: t − 1 ht− 1 . The value ft is between 0 and 1. Then the input gate calculates
( ( ) ) it and C
̃ t updates the cell state in the current cell to cooperate with Tanh
xlj = f αlj down xl−j 1 + blj (4) layer. Subsequently, output ht and cell state Ct in state t is generated in
the output gate. The process of the LSTM network can be depicted as
where xl−i 1 denotes the j th output feature map of the (l − 1) layer. follows [62]:
down( ⋅) is the max-pooling subsampling function and blj is the bias. After ( )
ft = σf Wf ⋅ [ht− 1 , xt ] + bf (6)
the convolution and max-pooling processes are completed, the output
feature maps will contain local correlation information among multiple it = σi (Wi ⋅ [ht− 1 , xt ] + bi ) (7)
energy resources.
ot = σ o (Wo ⋅ [ht− 1 , xt ] + bo ) (8)
2.3.2. Nonlinear temporal pattern extraction
The deep learning framework of A-LSTM for nonlinear temporal ̃ t = tanh(Wc ⋅ [ht− 1 , xt ] + bc )
C (9)
pattern extraction is shown in Fig. 4. The power generation of renewable
energy resources and weather conditions represent strong time-varying ̃t
Ct = ft ⋅Ct− 1 + it ⋅C (10)
characteristics. The LSTM network is applied to approximate the
nonlinear relationship between the variables in the previous period and ht = ot ⋅tanh(Ct ) (11)
the current state to extract nonlinear temporal patterns. The energy
power generation and corresponding weather conditions are con­ where (Wf , bf ), (Wi , bi ) and (Wo , bo ) represent the weights and biases of
structed as a two-dimensional matrix xlstm in Eq. (5), of which the col­ three control gates, and σ( ⋅) is the sigmoid activation function.
umns represent weather conditions and single energy generation, and As shown in Fig. 4, the conventional LSTM cannot take the impor­
the rows denote different time states. tance of different time states into account. It is necessary to consider the
association between the output and corresponding input after obtaining
xlstm = (w1 , w2 , ..., wn , yi )m m ∈ [T − a, T − 1] (5)
the hidden states of LSTM [63] to overcome this drawback. The atten­
where wn denotes weather condition variable and yi is one of the mul­ tion mechanism is applied to assign different weight features to the
correlative parts of inputs. In this way, the LSTM network can focus on
the significant time states which possess a great influence on the con­
cerning output. The formula of the attention mechanism can be illus­
trated as follows [64]:
ei = u tanh(whi + b) i = 1, 2, ...., t (12)

exp(ei )
αi = ∑ (13)
i exp(ei )

∑
C= ai hi (14)
i

where b represents the bias u and w denote the weighting coefficients of

the attention layer. ei is the probability distribution of the attention layer
for the hidden state hi . αi is the weight of each hidden state, and C is the
weighted feature.

2.3.3. Linear temporal pattern extraction

In general, the AR structure is commonly applied for stationary data.
In this work, the AR model is established to extract the linear temporal
Fig. 4. The basic architecture of the A-LSTM network.

6
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

pattern of multiple energy resources. In the process of linear pattern evaluated by applying a real-world case of RES. The energy resources in
extraction, yt depending just on its (a − 1) previous values, the general the selected case of RES include solar PV, solar thermal, and wind. To
equation on the AR model can be expressed as follows: ensure the normal operation of the renewable energy system, it is crucial
to predict power generation accurately to reduce the supply risk of
yt = b + w0 yt− 1 + ... + wa− 2 yt− + εt (15)
a+1
brownouts in the system and lower costs for balancing energy. To
where b is the intercept, and w0 , w1 , ..., wa− 2 are the coefficient of re­ illustrate the superiority of the proposed hybrid model, several
gressors. εt is white noise that obeys the normal distribution, and the advanced prediction models, including extreme gradient boost (XGB)
meaning is zero. and ANN, are utilised for comparison. Furthermore, the sensitivity
analysis of different component combinations is carried out to explain
2.3.4. Power generation prediction via the hybrid framework the significance of different information patterns.
These three different functional modules separately extract different
information from the input data. After acquiring various data patterns, 3.1. Data description
the outputs of functional modules are concatenated. Hence, the fusion
results contain three different information, namely, energy correlation This study is designed for the hourly power generation of multiple
pattern, nonlinear temporal pattern, and linear temporal pattern. Sub­ renewable energy resources for RES. It is important for reliable and
sequently, a fully connected layer is employed to reduce the feature efficient model development employing a comprehensive dataset. As a
dimensions and output the final predicted results. The formulations consequence, the power generation data from a real RES are collected,
between Eq (2) to Eq (15) can be expressed as Y = NN(X; θ). Then the which includes the measured power generation of multiple renewable
predicted errors called mean squared error (MSE) of the hybrid frame­ energy resources. The collected data are from November 1, 2016, to
work can be represented as follows: October 31, 2017, with a 1 h resolution. It is necessary to conduct data
division to ensure the robustness and stability of the proposed hybrid
1 ∑ framework. In this study, three sets of power generation data are
N
L(θ) = MSE = |NN(xi ; θ) − yi |2 (16)
N i=1 divided; namely, 70% of datasets for model training, 20% of datasets for
model testifying after each epoch is completed, and 10% of datasets to
where xi is the input data, and yi is the observed value of power gen­ verify the accuracy and efficiency of the prediction models.
eration.
3.2. Experimental setting
2.4. Evaluation metrics
The hybrid framework and other deep learning models are built
Several evaluation indices are used to evaluate the accuracy and using PyTorch, which supports the dynamic network structure. The
efficiency of the proposed hybrid prediction framework. The commonly Adam optimiser and MSE loss function are employed to update network
used root mean squared error (RMSE), mean absolute error (MAE), and parameters to optimise the training process of the deep learning
mean absolute percentage error (MAPE) are applied to reflect the de­ framework, and the training epoch is set as 100 because the network is
viation between the predicted values and the observed values. These deep enough, and the initial learning rate is set as 0.001. The early
three indices can be expressed as follows: stopping method is employed, and the number of iterations is set as 20 to
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ avoid the over-fitting of prediction models. For the establishment of the
√1 ∑ deep learning network, the model parameters play a significant role in
N
RMSE = √ (yi − ̂ y i )2 (17)
N i=1 predictive performance. Therefore, it is necessary to determine a suit­
able model structure. The parameters of the hybrid model contain three
sections named the CNN network parameters, the LSTM network pa­
1 ∑N
MAE = |yi − ̂y i | (18) rameters, and the fully connected network parameters. The parameters
N i=1
of the CNN network include convolution layer, kernel size, convolution
n ⃒ ⃒ stride, pooling size, and pooling stride. As depicted in Table 1, the range
1∑ ⃒yi − ̂
y i ⃒⃒
MAPE = ⃒ × 100% (19) is 1–5, 1–4, 16–256, 0–0.5, and 64–512 for strides, size, number of
n i=1 ⃒ yi ⃒ layers, filters, dropout, and batch size. By means of trial and error, the
Additionally, to testify to the efficiency and reliability of the pre­ suitable parameters of the CNN network can be selected, and the
diction models, the R2 score and the confidence index (CI) are employed determined results are shown in Table 1.
for further comparison. The mathematical formulation of these two The parameters of the LSTM network include time step, neural units,
metrics is as follows: batch size, number of layers, and dropout. After training the hybrid
models several times, the suitable parameters can be determined. As
∑
N
depicted in Table 2, the number of layers is set as 2, and the neural units
(yi − ̂y i )2
are 200 and 150. As depicted in Fig. 5, there is a peak in the 24 h time
2
R =1− i=1
(20)
∑N length for all three renewable energy resources, which indicates that the
(yi − yi )2
i=1 patterns of these three types of energy resources are the most repetitive
⎛ ∑n ⎞⎛ ∑n ⎞
y i )2
(yi − ̂ y i )2
(yi − ̂ Table 1
⎜ ⎟ ⎜ ⎟ The parameters in the CNN network.
CI = ⎜ i=1
⎝1 − ∑n
⎟⋅⎜1 −
⎠⎝ ∑
n
i=1 ⎟
⎠ (21)
2 2
(yi − yi ) (|yi − yi | + |̂
y i − yi |) Parameters Range Result
i=1 i=1
Strides (conv layer) 1–5 1
Strides (pooling layer) 1–5 1
where N is the total number of data, yi , ̂
y i , and yi represent the original
Kernel size 1–4 2
values, the predicted results, and the average values of original values. Pooling size 1–4 2
Number of layers 1–5 3
3. Case study Filters 16–256 [40, 80, 160]
Dropout 0–0.5 0.3
Batch size 64–512 64
In this section, the proposed hybrid framework is extensively

7
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Table 2 with a 24 h interval. As a consequence, the time step is selected as 24.

The parameters in the LSTM network. The batch size and the dropout are similar to the CNN network.
Parameters Range Result Additionally, the parameters in the fully connected layer contain the
batch size, neural units, and a number of layers. As shown in Table 3, the
Time step 2–30 24
Neural units 1–300 [200, 150] process of trial and error is conducted to select the suitable network
Batch size 64–512 64 structure. The number of network layers is selected as 2, and the neural
Number of layers 1–3 2 units are set as [100, 10]. The dropout and the batch size are similar to
Dropout 0–0.5 0.3 the CNN network.

Fig. 5. The autocorrelation coefficient of multiple energy resources (a) Solar thermal (b) Wind total (c) Solar PV.

8
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Table 3 Table 4
The parameters in the fully connected network. The MAPEs of the different models on multiple energy sources.
Parameters Range Result Renewable XGB ANN A- A–CNN–LSTM Hybrid
resources LSTM framework
Neural units 1–300 [100, 10]
Batch size 64–512 64 Solar PV 67.22% 52.14% 32.98% 19.75% 9.16%
Number of layers 1–3 2 Solar thermal 50.22% 46.56% 35.8% 27.66% 18.15%
Dropout 0–0.5 0.3 Wind power 41.58% 35.45% 31.68% 26.56% 16.87%

3.3. Prediction results of power generation obvious that the prediction errors of the hybrid framework are smaller
than other advanced prediction models, with MAEs being 179.47 kW,
After determining the suitable network structure, the power gener­ 23.86 kW, and 130.19 kW in all cases. The hybrid framework acquires
ation prediction model is constructed. Several advanced prediction lower predicted errors and achieves a 53.6% reduction of MAPEs for
models are applied for performance comparisons, such as XGB, ANN, A- solar PV compared with A–CNN–LSTM. The lowest prediction errors
LSTM, and attention-based convolution neural network long-short term also indicate a better generalisation ability for the proposed hybrid
memory (A–CNN–LSTM), aiming to embody the effectiveness of the framework. As for the conventional single models, the MAEs of XGB and
proposed hybrid framework. Owing to the traditional application, the ANN keep at a high level, which may result from the absence of
input of other advanced prediction models is information 2, which only comprehensive information extraction ability. On the other hand, the
includes weather conditions and a single energy resource. Based on the MAEs and RMSEs of XGB and ANN are not robust in all the case studies,
evaluation metrics mentioned in the previous content, the prediction which is reflected by the large variation range of metrics.
errors of different models can be depicted in Fig. 6 and Table 4. It is The A-LSTM deep learning network can obtain better prediction
performance by combining the memory block and attention block. The
MAEs of A-LSTM for solar PV, solar thermal, and wind power are 246.21
kW, 30.16 kW, and 150.41 kW. By comparing the errors of A-LSTM and
A–CNN–LSTM, it can be seen that the model that has a better fitting
ability of local dependency between variables will possess better per­
formance. Combining the local dependency between variables and long-
term dependency will lead to a better prediction ability, and the pre­
diction errors of A–CNN–LSTM are lower than A-LSTM, with MAEs being
235.42 kW, 28.26 kW, and 141.54 kW. Different from other traditional
prediction methods, the hybrid model takes energy correlation patterns,
and nonlinear and linear temporal patterns into consideration. As such,
with comprehensive data information support, the prediction accuracy
and efficiency of the hybrid framework are greatly improved. The pre­
diction accuracy of the hybrid framework raises sharply compared to
A–CNN–LSTM, its MAEs reduced by 55.95 kW, 4.4 kW, and 11.35 kW for
solar PV, solar thermal, and wind power. Accordingly, using deep
learning frameworks to extract different pattern information of power
generation for multiple renewable resources is a practical and effective
approach.
The power generation values predicted by different models are
depicted in Fig. 7 to compare the prediction results more graphically.
Overall, the predicted results of the hybrid framework are closer to the
observed values in all cases than other prediction models, which in­
dicates the proposed hybrid framework possesses a better prediction
ability and has a higher degree to the observed values. In each case, the
prediction accuracy of A–CNN–LSTM is a little lower but close to that of
the hybrid framework.
In the solar PV and solar thermal cases, the prediction curves of ANN
and XGB have a larger deviation than that of other prediction models.
Particularly, during the period between sunrise and sunset, the pre­
dicted curves of XGB and ANN deviate greatly from the observed values
of power generation. Furthermore, after sunset, the predicted values of
XGB and ANN are negative. This similar situation also happens in the
predicted curves of A-LSTM and A–CNN–LSTM in the solar PV case.
However, the predicted curves of the proposed hybrid framework not
only follow a large portion of the observed values but also obey the
scientific theory in all these two cases. The predicted values of the
hybrid framework after sunset are zero. In the wind power case, the
observed power generation fluctuates in all the periods, which makes
the prediction of power generation for multiple renewable resources
more difficult. As shown in Fig. 7 (c), the deviation between the pre­
diction curves of prediction models and the observed curve is relatively
big. In response, comparing the results of A-LSTM and the hybrid
framework, it can be seen that the A-LSTM outputs the worse results due
Fig. 6. Prediction errors of different models on multiple energy resources (a)
to the neglect of the energy correlation pattern and linear temporal
Solar PV (b) Solar thermal (c) Wind power.

9
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Fig. 8. The CIs of prediction models on different energy resources.

contrast, CI values of the XGB model are the lowest among other models,
which indicates that the XGB model is not suitable for the underlying
problem. Comparing the A-LSTM model and the A–CNN–LSTM model, it
is proved that the modelling of local dependency between variables
brings better accuracy and reliability to the prediction model.
Fig. 9 further illustrates the effective performance of the proposed
hybrid framework, which depicts the scatter plots of different prediction
models. By showing the scatter plots, the linear relationship between
predicted power generation and observed values is displayed intuitively.
The scatter plots represent the accurately estimated cumulative value of
power generation near the unit slope line (black line) for different pre­
diction models, and the R2 score is depicted in the upper left of the
figures. For conventional machine learning models, such as ANN and
XGB, the scatter plots are further to the observed power generation
values due to the poor extraction ability of temporal patterns. The A-
LSTM network utilises memory block and attention block to extract
temporal information, which leads to closer results than ANN and XGB.
With the assistance of a convolution block, the A–CNN–LSTM network
models the local dependency between variables, which results in more
accurate results than A-LSTM. However, these models neglect the local
correlation between multiple energy resources and linear temporal
patterns. The proposed hybrid framework applies different deep
learning networks to extract comprehensive information from input
data, and the predicted results are closer to observed values than other
advanced models.
By comparing the proposed model with other comparative models,
the accuracy and efficiency of the proposed model are illustrated. For a
real-world RES, the historical power generation data are collected and
pre-processed to train the proposed model. After acquiring the model
parameters, the future weather conditions are gathered to input into the
trained model. The future power generation of multiple renewable en­
Fig. 7. Prediction results of models on multiple renewable energy sources (a) ergy resources can be obtained.
Solar PV (b) Solar thermal (c) Wind power.

pattern. When applying the CNN network to model the local dependency 3.4. Sensitivity analysis
between variables, the results of A–CNN–LSTM are more accurate than
A-LSTM. The proposed framework extracts data features from three The performance comparisons of different prediction models are
different information, an energy correlation feature from information 1, carried out in section 3.2, and the results suggest that the proposed
a nonlinear temporal feature from information 2, and a linear temporal hybrid framework outperforms other advanced prediction models. To
feature from information 3. As such, the prediction results of the hybrid further explain the significance of each different information pattern,
framework are more accurate than other advanced models. sensitivity analysis of different components combination is conducted.
CI values of all prediction models are explored as well to provide a The hybrid framework includes three different information patterns
more robust assessment of the model performance. The CI can give a which are energy correlation patterns, nonlinear temporal patterns, and
comprehensive assessment of model accuracy and reliability. As depic­ linear temporal patterns. The comparative models will leave out any
ted in Fig. 8, it is further ascertained that the hybrid framework pos­ kind of information pattern. Therefore, three comparative models, the
sesses better accuracy and reliability as higher CI values are seen in model without information 1 (A-LSTM-AR), the model without infor­
comparison to the comparative A–CNN–LSTM, A-LSTM, ANN, and XGB. mation 2 (CNN-AR), and the model without information 3
The hybrid framework outputs the most accurate and reliable results (A–CNN–LSTM), can be established.
among all the prediction models examined in this study for hourly power Multiple experiments are conducted to test the prediction errors of
generation prediction of multiple renewable energy resources. In different models. Fig. 10 shows that RMSEs and MAEs of different pre­
diction models on multiple renewable resources. Overall, the hybrid

10
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Fig. 9. Scatter plots presentations of prediction models on different energy sources (a) Solar PV (b) Solar thermal (c) Wind power.

11
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

Fig. 10. The prediction errors of different components combination (a) Solar PV (b) Solar thermal (c) Wind power.

framework, which includes all three types of information patterns, A–CNN–LSTM, the prediction errors of RMSE are close to each other. In
outperforms other comparative models and has the lowest prediction the cases of solar thermal and wind power, the maximum of RMSEs and
errors. As shown in Fig. 10 (a), due to the neglect of one kind of infor­ MAEs of A–CNN–LSTM is lower than that of CNN-AR. However, in the
mation pattern, the prediction accuracy of comparative models is case of solar PV, the maximum prediction errors of A–CNN–LSTM are
decreased. A similar issue also happens in the other two energy re­ higher than that of CNN-AR. But overall, the prediction performance of
sources. In all three cases, it can be found that the A-LSTM-AR network A–CNN–LSTM is better than CNN-AR. Consequently, it is crucial for a
depicts the worst stability, and the prediction errors are the highest. more accurate and robust power generation prediction of multiple
Comparing the hybrid framework and A-LSTM-AR, the prediction per­ renewable energy sources to extract comprehensive data patterns
formance is worse if the model only takes the nonlinear and linear simultaneously. It should be noted that the A–CNN–LSTM constructed in
temporal patterns into consideration. It is also proved that there exist this subsection differs from the model in section 3.2.
correlation interactions between multiple renewable resources, and the
energy correlation possesses great significance on performance 4. Conclusion
improvement. The prediction errors of the hybrid framework also depict
lower fluctuation than that of A-LSTM-AR, which indicates that the Renewable energy has developed rapidly and has received increasing
robust prediction performance of the hybrid framework is better than A- attention recently. For a renewable energy system, accurate short-term
LSTM-AR. Comparing A–CNN–LSTM to A-LSTM-AR and CNN-AR, it can prediction for power generation is crucial for ensuring energy supply
be seen that the lack of the energy correlation pattern has the greatest and demand balance and decreasing risks. This paper aims to develop an
impact on model performance. For the temporal information pattern, accurate and efficient model for power generation prediction of multiple
the nonlinear temporal pattern has higher importance than the linear renewable resources to fill the gaps of ignoring the relationship among
temporal pattern from the comparison between A-LSTM-AR and CNN- different energy resources. Accordingly, a novel hybrid framework is
AR. Additionally, the prediction errors of the A-LSTM-AR model repre­ established, and it takes energy correlation patterns, nonlinear temporal
sent the highest fluctuation among prediction models. The fluctuation of patterns, and linear temporal patterns into account. The local correla­
prediction errors output by A–CNN–LSTM is a little lower but close to tions between different energy sources are first introduced in this study
that of A-LSTM-AR. Comparing the hybrid framework and to assist the model development. A real renewable energy system is

12
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

taken as an example, and a comparison with several advanced models is References

performed to evaluate the effectiveness of the proposed hybrid frame­
work. The results indicate that the hybrid framework outperforms other [1] Yan J, Yang Y, Elia Campana P, He J. City-level analysis of subsidy-free solar
photovoltaic electricity price, profits and grid parity in China. Nat Energy 2019;4:
advanced models, with RMSEs being 347.549 kW, 49.856 kW, 176.784 709–17.
kW, and R2 higher than 0.945 in all cases. Subsequently, a sensitivity [2] Wang C, Zheng J, Du J, Wang G, Klemeš JJ, Wang B, et al. Weather condition-based
analysis of different component combinations is carried out, and the hybrid models for multiple air pollutants forecasting and minimisation. J Clean
Prod 2022;352:131610.
results suggest that the model extracting comprehensive information [3] Du J, Zheng J, Liang Y, Lu X, Klemeš JJ, Varbanov PS, et al. A hybrid deep learning
patterns has better accuracy and robustness than other comparative framework for predicting daily natural gas consumption. Energy 2022:124689.
models. The prediction errors of the hybrid framework are lower than [4] Wang Y. The analysis of the impacts of energy consumption on environment and
public health in China. Energy 2010;35:4473–9.
other models, with MAEs from 47.745 kW, 347.549 kW, and 175.765 [5] Madurai Elavarasan R, Afridhis S, Vijayaraghavan RR, Subramaniam U,
kW, to 51.217 kW, 394.42 kW, and 181.678 kW, in all three cases. For Nurunnabi M. SWOT analysis: a framework for comprehensive evaluation of
the power generation prediction of multiple renewable energy, the en­ drivers and barriers for renewable energy development in significant countries.
Energy Rep 2020;6:1838–64.
ergy correlation possesses higher significance than the other two types
[6] Lerman LV, Benitez GB, Gerstlberger W, Rodrigues VP, Frank AG. Sustainable
of information patterns. Additionally, nonlinear temporal information is conditions for the development of renewable energy systems: a triple bottom line
more important than linear temporal information for model perspective. Sustain Cities Soc 2021;75:103362.
development. [7] IEA. World. Energy outlook 2016, executive summary. International Energy
Agency; 2016. p. 1–8.
In this work, the local correlation patterns between different energy [8] Aslam S, Herodotou H, Mohsin SM, Javaid N, Ashraf N, Aslam S. A survey on deep
resources in a single renewable system are considered, and the predic­ learning methods for power load and renewable energy forecasting in smart
tion performance is greatly improved. However, the regions’ de­ microgrids. Renew Sustain Energy Rev 2021;144:110992.
[9] Patel AM, Singal SK. Optimal component selection of integrated renewable energy
pendency patterns between different renewable systems are neglected, system for power generation in stand-alone applications. Energy 2019;175:
leading to a limited generalisation ability on different RESs. To further 481–504.
improve the model accuracy and generalisation ability, in our future [10] Elliston B, MacGill I. The potential role of forecasting for integrating solar
generation into the Australian national electricity market. Proceedings of the Solar
work, the region’s correlations between different renewable energy 2010.
systems on the performance of the prediction model is going to be [11] Severiano CA, Silva PCdLe, Weiss Cohen M, Guimarães FG. Evolving fuzzy time
further studied. series for spatio-temporal forecasting in renewable energy systems. Renew Energy
2021;171:764–83.
[12] Liew PY, Varbanov PS, Foley A, Klemeš JJ. Smart energy management and
Credit author statement recovery towards Sustainable Energy System Optimisation with bio-based
renewable energy. Renew Sustain Energy Rev 2021;135:110385.
[13] Naderipour A, Abdul-Malek Z, Nowdeh SA, Kamyab H, Ramtin AR, Shahrokhi S,
Jianqin Zheng: Conceptualisation, Methodology, Investigation,
et al. Comparative evaluation of hybrid photovoltaic, wind, tidal and fuel cell clean
Writing - Original Draft; Jian Du: Methodology, Software, Writing - system design for different regions with remote application considering cost.
Original Draft; Bohong Wang: Conceptualisation, Methodology, J Clean Prod 2021;283:124207.
Writing - Review & Editing; Jiří Jaromír Klemeš: Funding acquisition, [14] Pan X, Uddin MK, Han C, Pan X. Dynamics of financial development, trade
openness, technological innovation and energy intensity: evidence from
Writing - Review & Editing, Proofreading; Qi Liao: Writing - Review & Bangladesh. Energy 2019;171:456–64.
Editing; Yongtu Liang: Supervision, Funding acquisition, Writing - [15] Cergibozan R. Renewable energy sources as a solution for energy security risk:
Review & Editing empirical evidence from OECD countries. Renew Energy 2022;183:617–26.
[16] Wang G, Wang F, Shen F, Chen Z, Hu P. Novel design and thermodynamic analysis
of a solar concentration PV and thermal combined system based on compact linear
Declaration of competing interest Fresnel reflector. Energy 2019;180:133–48.
[17] Law EW, Kay M, Taylor RA. Evaluating the benefits of using short-term direct
normal irradiance forecasts to operate a concentrated solar thermal plant. Sol
The authors declare that they have no known competing financial Energy 2016;140:93–108.
interests or personal relationships that could have appeared to influence [18] Kabalci E. Design and analysis of a hybrid renewable energy plant with solar and
the work reported in this paper. wind power. Energy Convers Manag 2013;72:51–9.
[19] Zendehboudi A, Baseer MA, Saidur R. Application of support vector machine
models for forecasting solar and wind energy resources: a review. J Clean Prod
Data availability 2018;199:272–85.
[20] Wang F, Xuan Z, Zhen Z, Li Y, Li K, Zhao L, et al. A minutely solar irradiance
forecasting method based on real-time sky image-irradiance mapping model.
Data will be made available on request.
Energy Convers Manag 2020;220:113075.
[21] El Alani O, Abraim M, Ghennioui H, Ghennioui A, Ikenbi I, Dahr F-E. Short term
Acknowledgement solar irradiance forecasting using sky images based on a hybrid CNN–MLP model.
Energy Rep 2021;7:888–900.
[22] Hoyos-Gómez LS, Ruiz-Muñoz JF, Ruiz-Mendoza BJ. Short-term forecasting of
This work was part of the Program “Study on Optimisation and global solar irradiance in tropical environments with incomplete data. Appl Energy
Supply-side Reliability of Oil Product Supply Chain Logistics System” 2021:118192.
funded under the National Natural Science Foundation of China, grant [23] Huang X, Li Q, Tai Y, Chen Z, Zhang J, Shi J, et al. Hybrid deep neural model for
hourly solar irradiance forecasting. Renew Energy 2021;171:1041–60.
number 51874325. Additional funding comes from EU project “Sus­ [24] Kumari P, Toshniwal D. Deep learning models for solar irradiance forecasting: a
tainable Process Integration Laboratory – SPIL”, project No. CZ.02.1.01/ comprehensive review. J Clean Prod 2021;318:128566.
0.0/0.0/15_003/0000456 funded by EU “CZ Operational Programme [25] Reikard G. Predicting solar radiation at high resolutions: a comparison of time
series forecasts. Sol Energy 2009;83:342–9.
Research, Development and Education”, Priority 1: Strengthening ca­ [26] Zeng J, Qiao W. Short-term solar power prediction using a support vector machine.
pacity for quality research, innovation programme “Renewable energy Renew Energy 2013;52:118–27.
system for residential building heating and electricity production – [27] Yang D, Dong Z. Operational photovoltaics power forecasting using seasonal time
series ensemble. Sol Energy 2018;166:529–41.
RESHeat”, Grant Agreement # 956255), Science Foundation of Zhejiang [28] Hossain M, Mekhilef S, Danesh M, Olatomiwa L, Shamshirband S. Application of
Ocean University (11025092122), and Scientific Research Fund of extreme learning machine for short term output power forecasting of three grid-
Zhejiang Provincial Education Department (Y202250605). The authors connected PV systems. J Clean Prod 2017;167:395–405.
[29] Li P, Zhou K, Lu X, Yang S. A hybrid deep learning model for short-term PV power
are grateful to all study participants.
forecasting. Appl Energy 2020;259:114216.
[30] 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.

13
J. Zheng et al. Renewable and Sustainable Energy Reviews 172 (2023) 113046

[31] Ahmed R, Sreeram V, Mishra Y, Arif M. A review and evaluation of the state-of-the- [49] Feng C, Cui M, Hodge B-M, Zhang J. A data-driven multi-model methodology with
art in PV solar power forecasting: techniques and optimisation. Renew Sustain deep feature selection for short-term wind forecasting. Appl Energy 2017;190:
Energy Rev 2020;124:109792. 1245–57.
[32] Zheng J, Zhang H, Dai Y, Wang B, Zheng T, Liao Q, et al. Time series prediction for [50] Zhang C, Liu N, Ming J, Sharma A, Ma Q, Liu Z, et al. Development of a novel solar
output of multi-region solar power plants. Appl Energy 2020;257:114001. energy controllable Linear fresnel photoreactor (LFP) for high-efficiency
[33] Zhao W, Zhang H, Zheng J, Dai Y, Huang L, Shang W, et al. A point prediction photocatalytic wastewater treatment under actual weather. Water Res 2022;208:
method based automatic machine learning for day-ahead power output of multi- 117880.
region photovoltaic plants. Energy 2021;223:120026. [51] Garrido-Perez JM, Ordóñez C, Barriopedro D, García-Herrera R, Paredes D. Impact
[34] Wang J, Guo L, Zhang C, Song L, Duan J, Duan L. Thermal power forecasting of of weather regimes on wind power variability in western Europe. Appl Energy
solar power tower system by combining mechanism modeling and deep learning 2020;264:114731.
method. Energy 2020;208:118403. [52] Zheng J, Du J, Liang Y, Liao Q, Li Z, Zhang H, et al. Deeppipe: a semi-supervised
[35] Rodat S, Tantolin C, Le Pivert X, Lespinats S. Daily forecast of solar thermal energy learning for operating condition recognition of multi-product pipelines. Process Saf
production for heat storage management. J Clean Prod 2016;139:86–98. Environ Protect 2021;150:510–21.
[36] Law EW, Prasad AA, Kay M, Taylor RA. Direct normal irradiance forecasting and its [53] Zheng J, Du J, Liang Y, Wang C, Liao Q, Zhang H. Deeppipe: theory-guided LSTM
application to concentrated solar thermal output forecasting – a review. Sol Energy method for monitoring pressure after multi-product pipeline shutdown. Process Saf
2014;108:287–307. Environ Protect 2021;155:518–31.
[37] Madaeni SH, Sioshansi R, Denholm P. How thermal energy storage enhances the [54] LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W, et al.
economic viability of concentrating solar power. Proc IEEE 2012;100:335–47. Backpropagation applied to handwritten zip code recognition. Neural Comput
[38] Wang Y, Zou R, Liu F, Zhang L, Liu Q. A review of wind speed and wind power 1989;1:541–51.
forecasting with deep neural networks. Appl Energy 2021;304:117766. [55] Zang H, Liu L, Sun L, Cheng L, Wei Z, Sun G. Short-term global horizontal
[39] Yan J, Liu Y, Han S, Wang Y, Feng S. Reviews on uncertainty analysis of wind irradiance forecasting based on a hybrid CNN-LSTM model with spatiotemporal
power forecasting. Renew Sustain Energy Rev 2015;52:1322–30. correlations. Renew Energy 2020;160:26–41.
[40] James EP, Benjamin SG, Marquis M. Offshore wind speed estimates from a high- [56] Barzegar R, Aalami MT, Adamowski J. Coupling a hybrid CNN-LSTM deep learning
resolution rapidly updating numerical weather prediction model forecast dataset. model with a boundary corrected maximal overlap discrete wavelet transform for
Wind Energy 2018;21:264–84. multiscale lake water level forecasting. J Hydrol 2021;598:126196.
[41] Lei M, Shiyan L, Chuanwen J, Hongling L, Yan Z. A review on the forecasting of [57] Liu H, Mi X-w, Li Y-f. Wind speed forecasting method based on deep learning
wind speed and generated power. Renew Sustain Energy Rev 2009;13:915–20. strategy using empirical wavelet transform, long short term memory neural
[42] Zuluaga CD, Álvarez MA, Giraldo E. Short-term wind speed prediction based on network and Elman neural network. Energy Convers Manag 2018;156:498–514.
robust Kalman filtering: an experimental comparison. Appl Energy 2015;156: [58] Qiao D, Li P, Ma G, Qi X, Yan J, Ning D, et al. Realtime prediction of dynamic
321–30. mooring lines responses with LSTM neural network model. Ocean Eng 2021;219:
[43] Wang Y, Xie Z, Hu Q, Xiong S. Correlation aware multi-step ahead wind speed 108368.
forecasting with heteroscedastic multi-kernel learning. Energy Convers Manag [59] Li Y, Tong Z, Tong S, Westerdahl D. A data-driven interval forecasting model for
2018;163:384–406. building energy prediction using attention-based LSTM and fuzzy information
[44] Wang Y, Ma H, Wang D, Wang G, Wu J, Bian J, et al. A new method for wind speed granulation. Sustain Cities Soc 2022;76:103481.
forecasting based on copula theory. Environ Res 2018;160:365–71. [60] Shumway RH, Stoffer DS, Stoffer DS. Time series analysis and its applications. New
[45] Ouyang T, Zha X, Qin L. A combined multivariate model for wind power York: Springer; 2000.
prediction. Energy Convers Manag 2017;144:361–73. [61] Cowpertwait PS, Metcalfe AV. Introductory time series with R. Springer Science &
[46] Jahangir H, Golkar MA, Alhameli F, Mazouz A, Ahmadian A, Elkamel A. Short-term Business Media; 2009.
wind speed forecasting framework based on stacked denoising auto-encoders with [62] Zheng J, Liang Y, Xu N, Wang B, Zheng T, Li Z, et al. Deeppipe: a customized
rough ANN. Sustain Energy Technol Assessments 2020;38:100601. generative model for estimations of liquid pipeline leakage parameters. Comput
[47] Liu H, Mi X, Li Y. An experimental investigation of three new hybrid wind speed Chem Eng 2021;149:107290.
forecasting models using multi-decomposing strategy and ELM algorithm. Renew [63] Han Y, Fan C, Xu M, Geng Z, Zhong Y. Production capacity analysis and energy
Energy 2018;123:694–705. saving of complex chemical processes using LSTM based on attention mechanism.
[48] Hu J, Heng J, Tang J, Guo M. Research and application of a hybrid model based on Appl Therm Eng 2019;160:114072.
Meta learning strategy for wind power deterministic and probabilistic forecasting. [64] Xiang L, Wang P, Yang X, Hu A, Su H. Fault detection of wind turbine based on
Energy Convers Manag 2018;173:197–209. SCADA data analysis using CNN and LSTM with attention mechanism.
Measurement 2021;175:109094.

14