1 Introduction

The combination of excessive population growth, uncontrolled depletion of natural resources and global warming, which contributes to climate change, has heightened the likelihood of uncertainty in the natural environment and disrupted ecological balance. The surge in population growth is directly linked to an escalating demand for food and energy. The same reason can be attributed to the uncontrolled removal of natural resources and the transformation of forest lands into agricultural or industrial land. Increasing imports in the energy sector also trigger the price increase per liter of gasoline, diesel, coal and fossil fuels. For this reason; the search for an alternative energy source has become extremely important and extremely necessary to maintain the energy security of the population. Solar energy, wind energy, geothermal and hydro energy or flowing water are the most appropriate alternatives to the existing energy source. All these energy sources can be infinite and renewed. Hydroelectric is one of the main renewable energy sources due to its low cost and its potential to rapidly respond to the highest demand load on electricity and its potential [1]. In addition, hydroelectric is a very clean source of energy and uses only water. Water can also be used for other purposes after producing electrical energy. Hydroelectric is currently considered to be the largest renewable source of electricity in the world. Traditionally, it is a cheap and clean source of electricity. However, renewable energy often has various properties that cause uncertainties in renewable energy power systems. Therefore, the prediction of renewable energy is an important way to cope with this problem [2]. Machine learning is a data-oriented process used to create a smart result. The basic steps of machine education include data collection and preprocessing, feature selection and extraction, model selection and model verification.

Machine learning (ML) techniques contribute to the use of these accumulated data and become increasingly important in various fields of science, finance and industry [1, 3]. Several studies were conducted based on the forecasting of the dam’s inflow and the application of an operating reservoir model to calculate the generation of hydropower [4,5,6,7]. Dehghani et al. In their study, they combined the Grey Wolf Optimization (GWO) method with an adaptive neuro-fuzzy inference system (ANFIS) to predict hydropower generation. The results of this paper showed that the GWO-ANFIS method could satisfactorily predict hydropower generation, while ANFIS failed in nine input–output combinations [7]. Condemi et al. proposed a methodology for estimating hydropower capacity and feature summarization mechanisms (such as principal component analysis and feature grouping techniques) using machine learning (ML) regression techniques. They showed how ML regression techniques can be used to model the relationship between meteorological and climatic variables and hydroelectric generation using water tanks. In the study, MLP shows the best results in the problem with 0.2593 and 0.2128 TWH root mean square error and mean absolute error [8]. Another study was carried out by Al Rayess and Keskin, which presents the results of Machine Learning Techniques for short-term estimation of the amount of energy produced by General Circulation Models (GCMS) data by the Almus Dam and Hydroelectric Power Plant in Tokat, Turkey [9]. The results show that the correlation value of the gradient-enhanced tree model is equal to 0.717, indicating that the gradient-enhanced tree model is the most successful model for the available data. Li et al. in their study, they explore the potential of a GA-SVM (Genetic algorithms-support vector machine (SVM)) estimation model to accurately predict short-term power generation energy from small hydroelectric resources [10]. This article concludes that the GA-SVM model is a very potential candidate for estimating small hydroelectric (SHP) short-term power generation energy.

This article presents a study on the prediction of hydroelectric plant production using ML and a novel hybrid Genetic Grey Wolf Optimizer (GGWO)-based Convolutional Neural Network/Recurrent Neural Network (CNN/RNN)-Long Short-Term Memory (LSTM) regression techniques and an overview of the study is given Fig. 1. According to meteorological data, a hydroelectric power production (PP) is determined by the amount of electricity that can be produced while taking into account the level of water in its reservoir.

Fig. 1
figure 1

An overview of the study

Full size image

The article’s remaining sections are organized as follows: Chapter 2 describes the data collection process, the dataset used, the two main methods of prediction used: Method I (Machine Learning Algorithms) and Method II (CNN/RNN Hybrid Method) and evaluation metrics. In Chapter 3, the study focuses on the application of machine learning algorithms to predict hydroelectric power production. The feature selection process using PCA (Principal Component Analysis) and the performance metrics of various machine learning models are discussed. In addition proposed CNN/RNN hybrid method for power production prediction in this section. The feature selection using the GGWO (Grey Wolf Optimizer) method is explained, and the parameters of the deep learning models (CNN and LSTM) are detailed. The findings of the study from both Method I and Method II are presented and analyzed in Chapter 3. The evaluation metrics for each prediction approach are compared, and the strengths and weaknesses of each method are discussed in Chapter 5. Chapter 6 provides a comprehensive conclusion, summarizing the key findings of the study.

The suggested methods include contributions below:

  • The used dataset’s feature which has time series of the feature is selected with the Principal Component Analysis (PCA) method by using the minimum–maximum method in the normalization of the features,

  • In the study, the prediction process firstly ML techniques; SVR (Support Vector Regression), ELM (Extreme Learning Machine), RFR (Random Forest Regression), ANN (Artificial Neural Networks), WKNNR (Weighted K-Nearest Neighbor),

  • In the study, the data are also predicted with the help of the new proposed hybrid Genetic Grey Wolf Optimizer (GGW0)-based Convolutional Neural Network/Recurrent Neural Network (CNN/RNN)-Long Short-Term Memory (LSTM) regression techniques,

  • Within the scope of the study, correlation coefficient (r), coefficient of determination (also known as R-squared or R2), mean square error (MSE) and its rooted variant (RMSE), mean absolute error (MAE) and their normalized equivalents (RRMSE, RMAE)) values are obtained,

  • Comparative presentation of existing studies in the literature is provided.

2 Material and methods

This study predicts the electrical power production as well as the net head value representing the net water height of the Hirfanlı dam (this value is the primary component used for electricity production). In this part, dataset, data preprocessing, used algorithms and evaluation metrics are introduced.

2.1 Dataset

The final determinant of a region’s hydropower potential is typically related to water scarcity, which is in turn affected by precipitation and/or snowmelt, temperature and solar radiation. Therefore, it is evident that meteorological events have the greatest impact on the electrical generation of the hydroelectric power production (HEPP). However, there is no direct correlation between these variations in meteorological conditions and HEPP capacity. This indicates that it takes some time for precipitation to turn into water that can be used in the HEPP after it occurs [8]. Furthermore, the impact of meteorological events over time might not be consistent.

The Hirfanlı Dam-HEPP, a dam constructed for flood control, energy generation and irrigation purposes between 1953 and 1959, is chosen as the study’s sample plane. It is situated on the Kızılrmak River in Kaman/Kırşehir, Turkey. Hirfanlı HEPP, whose image and location in Turkey are shown in Fig. 2, is the second-largest HEPP in Kırşehir and produces 400 GWh of energy annually. It is located in Türkiye and has an installed capacity of 128 MWe. The total electrical power generation capacity of a HEPP is determined by the amount of electricity that can be produced based on the amount of water that is available and, consequently, the information on its net head, and the amount of water inlet originates from natural sources.

Fig. 2
figure 2

An overview of the study

Full size image

Starting from this perspective, Kaman Meteorology Directorate and Hirfanlı HEPP collected meteorological data and data on power production over a 15-year period, from 1 January 2007 to 31 December 2021, and the author organized the data to produce an original dataset. There are many different working variables because all meteorological and electrical generation data are determined daily. Two government organizations (Kaman/Kırşehir Weather Directorate and Hirfanlı Hydropower Station) have provided and verified these historical data. The dataset used in this study comprises valuable information from Table 1, encompassing essential meteorological features such as the highest, lowest and average air temperatures, as well as details regarding precipitation type, amount and intensity, spanning 15 years (365 days each year). In total, there are 64 variables captured daily over the entire period. This extensive dataset allows for a comprehensive analysis of factors influencing changes in water availability and electricity production in the region. By analyzing such a vast amount of data (350,400 data points in total), researchers gain a deeper understanding of the intricate relationships between meteorological variables and their impact on water resources and energy generation.

Table 1 Descriptive information of daily meteorological and power production parameters for Hirfanlı HEPP
Full size table

2.2 Data preprocessing

Data preprocessing tasks can significantly increase classification accuracy, depending on the domain and language, according to extensive experimental analyses. In addition to improving classification accuracy, it also has many other advantages like improved performance, smaller data sets and faster training cycles. When using data mining, it is not always appropriate to use raw data. When the value ranges of the variables in the data set differ from one another, normalization techniques are used. The pressure of variables with large mean and variance on other variables can be increased as a result of the variables’ significantly different means and variances, which can have an impact on accuracy and performance. To standardize the impact of each variable on the results, numerical data are normalized using normalization methods in data mining [11, 12]. Data are linearly normalized using the minimum-to-maximum approach. In other words, neither the magnitude nor the difference between the two values changes. The range of values after normalization is typically 0 to 1; minimum represents the lowest value that a data point can accept, and maximum represents the highest value that a data point can accept [13]. The minimum–maximum normalization procedure is carried out for the data used in this study during the data preprocessing stage. The values are derived from Eq. (1) for the minimum–maximum model.

$${\dot{x}} =\frac{{x}_{i}-{x}_{\min}}{{x}_{max}-{x}_{\min}}$$
(1)

\({\dot{x}}\): Normalized data, \({x}_{i}\): Input value, \({x}_{\max}\): Largest number in input, \({x}_{\min}\): Smallest number in the input.

2.3 Feature selection

Feature selection and Principal Component Analysis (PCA) are common techniques used in machine learning regression tasks to improve model performance and reduce the impact of overfitting. Feature selection involves selecting a subset of the most informative features from a dataset, while discarding redundant or irrelevant features. This can help to reduce the dimensionality of the data and improve the speed and accuracy of the model. PCA, on the other hand, involves transforming the original features into a new set of uncorrelated features, called principal components that capture the most variance in the data. This can also help to reduce the dimensionality of the data while retaining as much information as possible. A multivariate system with many variables that are related to one another is transformed by PCA into a system with new variables that, despite being less numerous and not related to one another as linear functions, can best explain the overall change of the prior system [14]. It is a method of multivariate statistical analysis. This technique transforms n features into m distinct features.

2.4 Algorithms

In essentially, classification is a technique, whereby certain judgments or predictions are made after learning and interpreting the information that is available, and the data are then divided into classes with labels based on those properties. There are many different fields, where the classification task is employed. Different algorithms are used by machine learning to solve issues. Various factors, including the nature of the problem, the quantity of variables and the kind of model, affect the algorithm that is employed [15,16,17]. The results of the study are obtained using 5 different machine learning algorithms, which are SVR (Support Vector Regression), ELM (Extreme Learning Machine), RFR (Random Forest Regression), ANN (Artificial Neural Networks), and WKNNR (Weighted K-Nearest Neighbor). The selected machine learning (ML) techniques (SVR, ELM, RFR, ANN, WKNNR) in the study are chosen based on their suitability for the prediction task of hydroelectric power production (PP) using meteorological data. Each technique has its strengths and advantages, making them valuable candidates for this specific regression problem.

2.5 Grid search

Machine learning models contain crucial parameters. Each method has its own set of parameters. These variables have a big impact on how well the model works. As a result, it is crucial to know the values of the parameters. The grid search method is one way to determine which parameter has what value so that the best model can be created. This approach involves processing each possible value for each parameter one at a time, calculating success and identifying the parameter values that are most appropriate. Hyperparameter fine-tuning can be done in the form of manual search, grid search, random search [18] and Bayesian optimization [19, 20]. The results of the grid search used to perform a thorough search and fine-tuning of all possible combinations of the specified hyperparameters are presented in this section.

2.6 Cross-validation

Cross-validation is a statistical technique used to assess the performance and generalization capability of a machine learning model [20]. It involves dividing the available data into multiple subsets, or “folds,” and using these folds for training and testing the model in a rotation [21]. The basic idea behind cross-validation is to ensure that the model’s performance is evaluated on different subsets of data to obtain a more reliable estimate of its effectiveness. In our study, classification is carried out using the tenfold cross-validation technique, and the results are reported.

2.7 Evaluation methods

Various metrics are utilized to assess the performance of predicting operations. Error values that are high indicate poor prediction accuracy. Some of the evaluation criteria that are used to determine the prediction error rate and accuracy are discussed in this section.

$$R=\frac{\sum ({x}_{i}-\overline{x })({y}_{i}-\overline{y })}{\sqrt{\sum {({x}_{i}-\overline{x })}^{2}\sum {({y}_{i}-\overline{y })}^{2}}}$$
(2)

R = correlation coefficient, \({x}_{i}\) = quantities of the variable of x, \(\overline{x }\) = mean of the quantities of x-variable, \({y}_{i}\) = values of the variable y-variable, \(\overline{y }\) = mean of the quantities of y-variable

$${\text{RMSE}}=\sqrt{\frac{1}{n}{\sum }_{i=1}^{n}{(\widehat{{y}_{i}}-{y}_{i})}^{2}}$$
(3)

RMSE = root mean squared error, n: number of data points, \(\widehat{{y}_{i}}\) = the predicted values, \({y}_{i}\) = the observed values

$${\text{MSE}}= \frac{1}{n}{\sum }_{i=1}^{n}{({y}_{i}-\widehat{{y}_{i}})}^{2}$$
(4)

MSE = mean squared error, n = number of data points, \(\widehat{{y}_{i}}\) = the predicted values, \({y}_{i}\) = the observed values

$${\text{MAE}}= \frac{{\sum }_{i=1}^{n}\left|{y}_{i}-\widehat{{y}_{i}}\right|}{n}$$
(5)

MAE = mean absolute error, n = number of data points, \(\widehat{{y}_{i}}\) = the predicted values, \({y}_{i}\) = the observed values.

3 Experimental results

The dataset obtained in this study is applied to machine learning (Method I) and deep learning (Method II) algorithms. For the Method I study, feature selection is performed for the features used in the dataset and analyzed in five different machine learning models. In the Method II, a novel GGWO and CNN/RNN model)-Long Short-Term Memory (LSTM) regression technique is introduced. Feature selection is done by using GGWO, and the proposed CNN/RNN-LSTM model is used for feature extraction and prediction of power production (PP).

3.1 Method I: prediction with machine learning algorithms

In this study, the net head value representing the net water height of the Hirfanlı dam (this value is the main component used for electricity generation) and the electrical power production are predicted. Within the scope of this study, the results are obtained by using machine learning and deep learning techniques separately. While 80% of the dataset is used for the training phase, 20% is used for the testing phase. For the test data, the data of 2011, 2016 and 2021 are chosen randomly. In this section, the feature matrix is reduced from the related variable in numbers 1 to 13 using the PCA method in Table 2. Input variables selected for PP prediction are given in Table 2. As histogram plots of the dataset, Fig. 3 also displays the frequency of some features.

Table 2 Selected features after using PCA method
Full size table
Fig. 3
figure 3

Distribution of the selected features of data for the input dataset as histogram plots (x-axis = VALUE × 1000)

Full size image

In this study, the grid search method is employed for hyperparameter tuning, which is a systematic approach to finding the optimal combination of hyperparameters for a machine learning model. The goal of hyperparameter tuning is to enhance the model’s performance by finding the parameter values that yield the best results on the validation set. Table 3 presents the specific parameter values used in the grid search. The grid search involves defining a set of possible values for each hyperparameter and then exhaustively searching through all possible combinations of these values. For each combination, the model is trained and evaluated using a predefined performance metric. By using the grid search method, researchers can efficiently explore a wide range of hyperparameter configurations to identify the optimal ones that lead to improved model performance.

Table 3 The hyperparameters of the used models
Full size table

Figure 4 displays the correlation between the features and the class label in the dataset. Upon examination, it becomes evident that there exists a strong correlation between the features NH, PWT and TOP, and the class label (PP). The high correlation observed between NH, PWT and TOP with the class label (PP) indicates that these features hold significant predictive power for determining the class label. In other words, variations in NH, PWT and TOP are closely associated with changes in the class label (PP), making them valuable indicators for the classification task. The results of ELM, which is the best classifier used in the study, are presented. The Power Production (GWh) vs Power Production prediction (GWh) graph for the ELM is given in Fig. 5. The importance of the features for the ELM model is shown in Fig. 6.

Fig. 4
figure 4

The correlation coefficient matrix of the predictor parameters for the daily meteorological data (The abbreviations of the parameter names are outlined in Table 2)

Full size image
Fig. 5
figure 5

Power Production (GWh) vs Power Production prediction (GWh) graph for the ELM model (inset: red line indicates the least square regression (y = mx + c), where y is the prediction and x is the real data, additionally r is the coefficient of the correlation)

Full size image
Fig. 6
figure 6

Importance of the features for ELM model prediction

Full size image

Within the scope of the study, correlation coefficient (r), coefficient of determination (also known as R-squared or R2), mean square error (MSE) and its rooted variant (RMSE), mean absolute error (MAE) and their normalized equivalents (RRMSE, RMAE)) values are obtained and the values obtained with the PCA method and without using the PCA method are given in Table 4. According to the MAE, MSE, RMSE, MAPE and R2 metrics of the five machine learning algorithm models tested for the prediction of Hirfanlı hydroelectric power production, the most successful result is obtained with the ELM algorithm model. When the ELM algorithm is used in the estimation process using the PCA method, the r, R2, RMSE and MAE values are obtained as 0.977, 0.955, 0.56 and 0.40, respectively, and they are the values with the best results. When Table 4 is examined, RMSE and MAE values are obtained as 0.56 and 0.4, respectively. In the literature, MSE and MAE values below 0.4 TWh are considered very good values and predictions over 1 TWh are regarded as being very poor [18]. The study highlights the importance of evaluating machine learning models using multiple metrics and comparing the results with actual values. The study also suggests that combining feature selection techniques such as PCA with machine learning algorithms can lead to improved prediction accuracy and reduced computational time.

Table 4 The testing performance metrics of the used ML algorithms
Full size table

3.2 Method II CNN/RNN hybrid method

3.2.1 Feature selection-based Grey Wolf Optimizer and genetic algorithm

Feature Selection (FS) is a popular and complex research area in various domains, such as machine learning, data mining and pattern recognition. It involves the identification of significant and informative attributes from a dataset, which is a crucial step in various applications, including predictive data analysis [22]. While FS is an attractive research topic, it also presents challenges that researchers must address to achieve accurate and efficient results. The authors have outlined three primary objectives of utilizing feature selection in classification [22]. Firstly, it aims to simplify the process of feature extraction, which can be a complex and time-consuming task. Secondly, it seeks to enhance the accuracy of classification models by selecting the most informative features that are relevant to the task at hand. Finally, feature selection is implemented to improve the reliability of performance estimation, ensuring that the classification model is producing consistent and dependable results. These objectives demonstrate the value and significance of feature selection in classification tasks across a range of domains.

Grey Wolf Optimizer (GWO): The proposed Grey Wolf Optimization algorithm imitates the social and hunting behavior of grey wolves. Grey wolves are divided into alpha, beta, delta and omega social groups. Because the wolf group obeys its rules, the alpha group is the dominant species. Secondary grey known as beta class support alpha in making decisions. The lowest ranking grey wolves are represented by Omega. A wolf is referred to as a delta if it does not fall under one of the aforementioned species. Along with their social interactions, grey wolves have an intriguing social behavior called group hunting. The phases of containing, hunting and attacking prey are the GWO’s main components [23, 24].

The feature selection problem involves selecting a subset of features to be used in a model [22]. To achieve this, researchers often represent potential subsets as binary strings of the same length as the total number of features. Each bit in the string indicates whether a corresponding feature is included or excluded from the subset. A value of ‘1’ indicates that the feature is selected, whereas a ‘0’ indicates exclusion. To model the social hierarchy of wolves, researchers typically use a system in which the most dominant wolves are considered alpha (α) and responsible for making decisions. Beta (β) wolves assist alpha in decision making and provide feedback. Delta (δ) wolves act as scouts and control the actions of omega (ω) wolves, which must obey all other wolves. In the GWO (Grey Wolf Optimizer) algorithm, alpha, beta and delta wolves lead the hunting process while omega wolves follow their lead [22]. The distance separating each wolf from the prey Eq. (6):

$$\overrightarrow{D}=\left|\overrightarrow{C.}\overrightarrow{{X}_{p}}(t)-\overrightarrow{X}(t)\right|$$
(6)
$$\overrightarrow{C}=2\overrightarrow{a}\overrightarrow{{r}_{2}}$$
(7)

where \(\overrightarrow{X}\) and \(\overrightarrow{{X}_{p}}\) denote the position vectors of the prey and a wolf, respectively, and t represents the current iteration. The coefficient vector \(\overrightarrow{C}\) is computed following the Eq. (7), and \(\overrightarrow{{r}_{2}}\) is a random vector with values between 0 and 1 [25].

$$\overrightarrow{X}\left(t+1\right)= \overrightarrow{{X}_{p}}\left(t\right)-\overrightarrow{A}.\overrightarrow{D}$$
(8)
$$\overrightarrow{A}=2\overrightarrow{a}\overrightarrow{{r}_{1}}-\overrightarrow{a}$$
(9)

Within the optimization process, \(\overrightarrow{A}\) represents a coefficient vector, and its elements, denoted as \(\overrightarrow{a}\) gradually decrease linearly from 2 to 0. Additionally, the vector \(\overrightarrow{{r}_{1}}\) is composed of random values falling within the interval [0, 1] and they are given in Eqs. (8) and (9). In the GWO algorithm, the alpha, beta and delta wolves act as leaders, guiding the omega wolves to move toward the optimal position [25]. To calculate the new position of a wolf mathematically, a specific formula is used.

$$\overrightarrow{X}\left(t+1\right)= \frac{{\overrightarrow{X}}_{1}+{\overrightarrow{X}}_{2}{+\overrightarrow{X}}_{3}}{3}$$
(10)

At a given iteration t in the GWO algorithm, the positions of the three best wolves in the swarm are identified and denoted as \(\overrightarrow{X}\) 1, \(\overrightarrow{X}\) 2 and \(\overrightarrow{X}\) 3. These wolves are used in a specific formula to determine the new position of each wolf in the pack. The movement pattern of wolves when |A| is less than 1 or greater than 1 is visually illustrated in Fig. 7.

Fig. 7
figure 7

The movement vectors’ behavior is influenced by the value of |A|. [25]

Full size image

The Genetic Algorithm considers the life adventure of living things. The good generations keep their lives, but the bad generations cannot save their lives and perish. The genetic algorithm acts within the framework of these principles. It is used in solving constrained problems, where there is no clear solution where the modeling process cannot be done mathematically [26]. In recent times, certain researchers have focused on solving the feature selection problem by employing genetic algorithms in a wrapper model approach. This method involves using genetic algorithms to search through all potential subsets of features in order to determine a set that can optimize the accuracy of a given learning algorithm for prediction [22]. By exploring the entire feature space, the wrapper model can identify the most informative and relevant features, leading to enhanced predictive performance. The use of genetic algorithms in feature selection highlights the importance of advanced techniques in data analysis and the potential for significant improvements in predictive modeling [22, 27].

In the study, the GGWO method suggested by Kihel and Chouraqui is used for feature selection [22]. This approach extends to metaheuristics combination, where various bio-inspired hybrid methods are utilized to solve feature selection problems. One such method is the GGWO, which combines the capabilities of GA and GWO algorithms to achieve optimal results. In the GGWO approach, the population of grey wolves is randomly initialized with either a bit 1 or 0, and each wolf’s fitness is evaluated using K-NN. The best, second-best and third-best solutions are designated as alpha, beta and delta, respectively. The position of the wolf is then updated by applying the crossover between X1, X2 and X3, resulting in improved feature selection performance. This hybridization approach demonstrates the potential for combining multiple techniques to enhance the efficiency and effectiveness of feature selection algorithms.

$$X(t+1)= {\text{Crossover}}({X}_{1},{X}_{2}{,X}_{3})$$
(11)

After initializing the population of grey wolves, their fitness is evaluated, and the positions of alpha, beta and delta are updated. The algorithm iterates until a stopping criterion is reached, which is typically the maximum number of iterations. Once the iterations are complete, the optimal feature subset is selected as the alpha solution. This process involves evaluating the fitness of each wolf and updating the positions of the alpha, beta and delta solutions in each iteration. By repeating this process, the algorithm can identify the most effective feature subset for the given problem. The use of the alpha solution as the optimal feature subset highlights the importance of selecting the best performing solution to achieve superior feature selection results. GGWO reduces the feature matrix from the closely related variable in numbers 1 to 9 in this section, and the features selected after using GGWO are given in Table 5. Table 6 shows the parameters which are selected by grid search methods.

Table 5 Selected features after using GGWO
Full size table
Table 6 The hyperparameters of deep learning models
Full size table

Convolutional Neural Networks (CNN): A CNN is a type of model that is used for image recognition, and is characterized by its ability to detect complex patterns in video frame images through nonlinearity [28]. By adjusting various hyperparameters, it becomes possible to represent complex patterns in a high-dimensional space. CNNs consist of several key components, including convolutional layers, pooling layers, a dense network and a softmax layer which generates the output [29]. Each of these components plays an important role in enabling the CNN to effectively recognize patterns within images. The proposed CNN parameters are detailed in Table 7. Input data in the form of time series of 2 dimensions of 16 cores, then maximum pooling is made. In the second, 32 core and maximum pools are made of 3-dimensional output tensors, the second evolution layer. Then the third evolution layer was 4 dimensions of 64 cores and maximum pool layer, then 128 core and maximum pool in the last evolution 5 dimensions. All maximum pooling layers received a maximum value with 2 dimensions. The data from CNN are applied to RNN in the form of a one-dimensional time series.

Table 7 CNN parameters
Full size table

Recurrent Neural Network (RNN): A Recurrent Neural Network (RNN) is a type of neural network that processes data sequentially, relying on its ability to retain a memory of the preceding sequence in order to learn effectively [30]. The name “recurrent” stems from the fact that the output from each time step is used as input for the subsequent step, building upon the output of the previous step by remembering it. By doing so, RNNs can capture long-term dependencies that exist within the training data.


Long Short-Term Memory (LSTM): LSTM, which is a variation of RNN, is utilized for generating captions from a range of video frames and images [28]. This is because LSTM is a highly effective sequence translation model that consists of numerous memory blocks known as cells. These cells are responsible for manipulating and retaining sequences using various categories of gates. One of the key benefits of using LSTM is that it helps overcome the vanishing gradient problem that can hinder the training of RNN models. As a result, LSTM has proven to be a highly useful tool for generating captions from visual data.

Figure 8 depicts the hybrid CNN/RNN model approach that has been put forward. This approach consists of four main components, as outlined in the diagram. Firstly, the feature vectors are generated by extracting features from video/images using a CNN and converting them into input dimensions. Secondly, a model is generated and used to provide a sequence through an RNN. Thirdly, semantic captioning is performed on the generated sequences. The evolution layer, LSTM recurrent layer and dense layer are the three types that make up the core layer. In order to extract properties, the input data sequence is first used as input into two subsequent evolution layers. The LSTM layer is then introduced to the resulting attribute map, and the three active LSTM gates are used to thoroughly learn about complex addictions. The gate of forgetting specifically removes irrelevant or unnecessary information from earlier cell conditions. The entrance gate stores the useful new information during the entrance.

Fig. 8
figure 8

Structure of the hybrid CNN/RNN model

Full size image

Additionally, signals from the cell status are filtered before being passed through the output passage and sent to the following state. The processed temporal data are used as input layers for non-linear transformation. The results of the predictions are then generated after the information obtained is reflected in the output space. Finally, the proposed model is evaluated using a variety of metrics, while also taking into account the data distribution and cross-validation. Figure 9 demonstrates Power Production (GWh) vs Power Production prediction (GWh) graph for the CNN/RNN hybrid model.

Fig. 9
figure 9

Power Production (GWh) vs Power Production prediction (GWh) graph for the CNN/RNN hybrid model (inset: red line indicates the least square regression (y = mx + c), where y is the prediction and x is the real data, additionally r is the coefficient of the correlation)

Full size image

According to the MAE, MSE, RMSE, MAPE and R2 metrics of proposed deep learning model tested for prediction of Hirfanlı hydroelectric power production, the most successful result is obtained with proposed deep learning model when comparing as Table 8. When proposed deep learning model is used in the prediction process and the r, R2 and MAE values are obtained as 0.9802, 0.961 and 0.314, respectively, and they are the values with the best results. Below are some examples of the literature studies. These studies in Table 9 are compared to the evaluation of the parameters compared to the models that we proposed better results than other methods.

Table 8 The testing performance metrics of the used ML and hybrid deep learning techniques
Full size table
Table 9 The testing performance metrics of the used ML and hybrid deep learning techniques
Full size table

4 Discussion

The objective of this study is to develop a predictive model for hydropower production using various machine learning and deep learning techniques. The research also delves into methodologies for constructing predictive models, underscoring the significance of feature selection approaches. Feature selection is crucial as it reduces the number of features, thereby minimizing learning time and improving prediction accuracy. The study utilizes meteorological data from the years 2007 to 2021, adjusting and normalizing outliers for estimating hydroelectric power plant production. The dataset is then split into 80% for training and 20% for testing. Model performance is assessed using metrics such as MAE, MSE, RMSE, MAPE and R2, with the PCA method employed in the analysis of algorithm results. According to metrics including MAE, MSE, RMSE, MAPE and R2, the ELM algorithm produced the best results among five different machine learning techniques for estimating hydroelectric capacity. Additionally, a novel technique is introduced, combining the Genetic Grey Wolf Optimizer (GGWO) with a Convolutional Neural Network/Recurrent Neural Network (CNN/RNN) model using Long Short-Term Memory (LSTM) for regression analysis. This approach involves utilizing GGWO for feature selection to reduce dataset dimensionality, followed by applying the CNN/RNN-LSTM model for feature extraction and prediction. The authors demonstrate the effectiveness of their proposed technique through experiments on real-world datasets, highlighting superior performance compared to other state-of-the-art techniques. Table 8 summarizes the results across all metrics for both machine learning algorithms and deep learning models. Future research avenues for this technique include exploring alternative optimization algorithms for feature selection besides GGWO. Additionally, investigating the impact of dataset size on the technique’s performance and exploring different architectures for the CNN/RNN-LSTM model could provide valuable insights. Moreover, extending the technique to incorporate other data types, such as time series or image data, and applying it to diverse domains like finance or healthcare would contribute to evaluating its performance in various applications.

5 Conclusion

This study focuses on predicting power production at the Hirfanlı dam using machine learning (ML) and deep learning (DL) techniques. Two distinct methods were employed: Method I utilized various ML algorithms with feature selection through the PCA method, while Method II introduced a CNN/RNN-LSTM hybrid model with feature selection based on the Grey Wolf Optimizer (GGWO) and Genetic Algorithm (GA). The study incorporated the tenfold cross-validation method and employed grid search for hyperparameter adjustments. In Method I, five ML models are utilized, with the Extreme Learning Machine (ELM) algorithm emerging as the most successful. It achieved a correlation coefficient (r) of 0.977 and a coefficient of determination (R2) of 0.955 with the best feature subset. Additionally, ELM demonstrated a mean absolute error (MAE) of 0.4, indicating high precision in power production prediction. Method II introduced the CNN/RNN-LSTM hybrid model, surpassing all ML models from Method I. This model demonstrated exceptional prediction accuracy with a correlation coefficient (r) of 0.9802 and a coefficient of determination (R2) of 0.961. Moreover, the CNN/RNN-LSTM model achieved a MAE of 0.314, highlighting its robustness and reliability in power production forecasting. In conclusion, this research provides valuable insights into renewable energy prediction, particularly in hydroelectric power production. ML and DL techniques, coupled with advanced feature selection methods, offer a robust framework for accurate and efficient power production forecasting. Future work should address factors such as data availability, model generalization across different dam systems, and the potential impact of external factors on power production. Additionally, exploring the integration of real-time data and considering the dynamic nature of environmental conditions could enhance the model’s predictive capabilities. These considerations are crucial for refining the models and ensuring their applicability in diverse operational scenarios. The outcomes of this study lay the foundation for future research in renewable energy prediction, inspiring the development of more sophisticated models for other energy sources and addressing the identified limitations.