The Effects of Increasing Ambient Temperature and Sea Surface Temperature Due to Global Warming on Combined Cycle Power Plant

Abstract

The critical consequence of climate change resulting from global warming is the increase in temperature. In combined cycle power plants (CCPPs), the Electric Power Output (PE) is affected by changes in both Ambient Temperature (AT) and Sea Surface Temperature (SST), particularly in plants utilizing seawater cooling systems. As AT increases, air density decreases, leading to a reduction in the mass of air absorbed by the gas turbine. This change alters the fuel–air mixture in the combustion chamber, resulting in decreased turbine power. Similarly, as SST increases, cooling efficiency declines, causing a loss of vacuum in the condenser. A lower vacuum reduces the steam expansion ratio, thereby decreasing the Steam Turbine Power Output. In this study, the effects of increases in these two parameters (AT and SST) due to global warming on the PE of CCPPs are investigated using various regression analysis techniques, Artificial Neural Networks (ANNs) and a hybrid model. The target variables are condenser vacuum (V), Steam Turbine Power Output (ST Power Output), and PE. The relationship of V with three input variables—SST, AT, and ST Power Output—was examined. ST Power Output was analyzed with four input variables: V, SST, AT, and relative humidity (RH). PE was analyzed with five input variables: V, SST, AT, RH, and atmospheric pressure (AP) using regression methods on an hourly basis. These models were compared based on the Coefficient of Determination (R²), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Square Error (MSE), and Root Mean Square Error (RMSE). The best results for V, ST Power Output, and PE were obtained using the hybrid (LightGBM + DNN) model, with MAE values of 0.00051, 1.0490, and 2.1942, respectively. As a result, a 1 °C increase in AT leads to a decrease of 4.04681 MWh in the total electricity production of the plant. Furthermore, it was determined that a 1 °C increase in SST leads to a vacuum loss of up to 0.001836 bar_a. Due to this vacuum loss, the steam turbine experiences a power loss of 0.6426 MWh. Considering other associated losses (such as generator efficiency loss due to cooling), the decreases in ST Power Output and PE are calculated as 0.7269 MWh and 0.7642 MWh, respectively. Consequently, the combined effect of a 1 °C increase in both AT and SST results in a 4.8110 MWh production loss in the CCPP. As a result of a 1 °C increase in both AT and SST due to global warming, if the lost energy is to be compensated by an average-efficiency natural gas power plant, an imported coal power plant, or a lignite power plant, then an additional 610 tCO₂e, 11,184 tCO₂e, and 19,913 tCO₂e of greenhouse gases, respectively, would be released into the atmosphere.

1. Introduction

The seriousness of the climate crisis, its devastating impacts, and the underlying causes are presented with evidence in the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report (AR6). The report assesses that climate change has led to an increase in extreme weather events, rising sea levels and temperatures, and the spread of droughts, floods, and wildfires. According to the report, from the period 1850–1900 to 2010–2019, the global surface temperature has increased by approximately 1.07 °C. This warming is largely attributed to greenhouse gas emissions resulting from human activities. The report also emphasizes that if current greenhouse gas emission levels persist, then it is likely that global temperature rise will exceed the 1.5 °C and 2 °C thresholds within the 21st century. Moreover, according to the report, the global average Sea Surface Temperature (SST) has increased by 0.88 °C (very likely range: 0.68–1.01 °C) since the beginning of the 20th century, and it is almost certain that the global ocean has been warming since at least 1971. Depending on greenhouse gas emissions and other climate change factors, SST is projected to increase by 1 to 3 °C by 2100 [1]. Accordingly, the reduction of emissions plays a crucial role in ensuring climate stability and promoting sustainability [2].

The Turkish State Meteorological Service (MGM) has published the “Turkey 2024 Climate Report”, which provides a comprehensive analysis of Turkey’s climate in 2024. According to the report, while Turkey’s long-term average temperature was approximately 13.5 °C, it increased to 15.5 °C in 2024. In the Marmara region, the long-term average temperature was around 14.5 °C, but it rose to approximately 16.5 °C in 2024 [3]. Additionally, according to the Official Climate Statistics prepared by the Turkish State Meteorological Service, a significant and continuous increase in the water temperature of the Marmara Sea has been observed between 1970 and 2024, with an annual rise of approximately 0.0454 °C. While the average temperature was around 14–15 °C in the 1970s, it has now exceeded 16 °C. This trend demonstrates the impact of global warming and climate change on the Marmara Sea [4]. As a result of all these developments, mitigating the impacts of climate change and securing the future of our planet has become an urgent priority today [5].

One effective method of utilizing energy resources more efficiently is electricity generation through combined cycle power plants (CCPPs). These power plants combine gas turbines and steam turbines within a dual thermodynamic cycle—typically the Brayton and Rankine cycles—to generate electricity in two sequential stages. In the first stage, natural gas is combusted in a gas turbine to produce high-temperature, high-pressure exhaust gases that drive the turbine and generate electricity. Rather than being released into the atmosphere, the hot exhaust gases are directed into a heat recovery steam generator (HRSG), where they are used to produce steam. This steam then powers a steam turbine, enabling a second stage of electricity generation without additional fuel consumption. The integration of these two cycles results in significantly higher overall thermal efficiency—often exceeding 60%—compared to conventional single-cycle power plants, which typically operate at efficiencies around 35–40%. This makes CCPPs a crucial technology for reducing fuel consumption and minimizing greenhouse gas emissions in modern power systems [6,7].

In a CCPP, Ambient Temperature (AT) has a significant impact on the plant’s efficiency and electricity generation. As AT increases, various negative effects can be observed in the plant’s operation. High AT leads to a reduction in the output power of the gas turbine. This, in turn, indirectly affects the steam turbine, resulting in less energy being produced by the turbines.

In wet-cooled power plants, the impact of AT primarily affects the load of gas turbines negatively, while its effect on steam turbines is relatively limited. In dry-cooled plants, in addition to the negative impact of air temperature on gas turbines, condenser vacuum deterioration also occurs, leading to more significant reductions in steam turbine efficiency. In hybrid cooling systems, although regional variations exist, they generally perform better than dry cooling systems but exhibit slightly lower performance compared to seawater cooling systems.

Similarly, in coal- or lignite-fired power plants, the efficiency of steam turbines varies depending on the type of cooling system used (wet, dry, or hybrid), as is the case in combined cycle power plants. As a result, on a global scale, rising air temperatures have the potential to reduce the overall efficiency of power plants and lead to losses in energy production.

Moreover, the main function of the condenser, which is one of the auxiliary equipment of the steam turbine, is to condense the steam and maintain the maximum vacuum level. Further expansion of the steam leads to a decrease in condenser pressure and an increase in vacuum. When energy losses are reduced, the net output power produced by the turbine increases. One of the most important factors affecting the reduction in condenser pressure is the cooling system. Lower cooling water temperatures result in lower pressure at the steam turbine output and, consequently, a higher vacuum. This increases the steam’s ability to perform work and enhances the overall electricity generation of the CCPP [8].

Additionally, the heating of stator and rotor windings in generators is directly related to the amount of power produced by the generator. Especially at high loads, the temperature of the generator windings increases significantly, which reduces the generator’s efficiency. Cooling systems in generators play a critical role in controlling winding temperatures. Effective cooling allows the generator to operate at higher power levels and prevents overheating. Depending on the generator’s design, water, air, or hydrogen cooling systems help maintain controlled winding temperatures, enabling sustained higher power production. Therefore, whether the cooling water temperature is low or high has a positive or negative impact on generator efficiency.

The performance of gas turbines is significantly influenced by environmental parameters, in addition to AT, such as RH and AP. An increase in RH facilitates the cooling of intake air in hot and dry climates, thereby improving gas turbine performance. In this context, systems such as evaporative coolers or chiller systems are used to humidify the intake air, which increases the air density, contributing to improved combustion efficiency and higher power output. AP is another critical factor affecting turbine performance. At higher altitudes, AP decreases due to reduced air density, which limits the mass flow of air entering the turbine and leads to a decline in energy production. In contrast, higher atmospheric pressure conditions enable more efficient operation of the compressor, resulting in a greater pressure differential across the turbine. This allows for increased mechanical work output and, consequently, improved overall thermal efficiency of the gas turbine cycle.

Since the increase in SST accelerates evaporation, it leads to a rise in RH. Moreover, in regions where SST is high, the heated air rises, leading to the formation of low-pressure systems at the surface. However, the connections of RH and AP to SST are not as strong as the relationship between AT and SST.

Prediction models developed to increase efficiency and reduce costs in the energy sector play a critical role in optimizing energy demand and production. While traditional regression analyses have been used in prediction models for many years, new methods such as Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), Decision Trees (DTs), and Gradient Boosting Regression provide much more accurate predictions. CCPP energy prediction models typically use the following data: The dependent variable (target variable) is Electrical Power Output (PE), while the independent variables (input variables) are AT, Exhaust Vacuum (V), Atmospheric Pressure (AP) and Relative Humidity (RH). Zhao and Foong developed an ANN model optimized using the Electrostatic Discharge Algorithm to predict PE in a CCPP. PE is predicted by considering the effects of AT, RH, AP, and V. The optimized ANN model using Electrostatic Discharge Algorithm outperformed the traditional ANN model in terms of both training and testing accuracy [9]. Elfaki and Ahmed examined the stochastic behavior of regression neurons, the effect of the number of neurons in hidden layers, the effect of data subset size for training, and the effect of the number of input variables over the ANN, which is being used for getting a regression model to predict PE of CCPP. The results showed that the ANN outperformed regression models in capturing nonlinear relationships, providing better accuracy [10]. Kaya et al. [11] compared local and global learning methods, utilizing regression techniques, neural networks, and clustering methods like K-means for predicting the PE of a CCPP. It has been determined that global models such as neural networks perform better in capturing complex relationships, while local models such as K-means excel in specific scenarios [11]. Santarisi and Faouri [12] aimed to predict the PE of a CCPP using regression algorithms such as Gradient Boosting Regression, Support Vector Regression, Ridge and Lasso regression, Multiple Linear Regression (MLR), and DT. The results showed that Gradient Boosting Regression delivered the best prediction performance [12]. Tüfekçi compared fifteen different Machine Learning (ML) regression methods to predict the hourly full-load PE of a CCPP. The most successful method was observed with Mean Absolute Error (MAE) of 2.818 and Root Mean Square Error (RMSE) of 3.787, using the REPTree with the Bagging algorithm [13]. Asghar et al. [14] used a Backpropagation Neural Network (BPNN) to predict the power output for investigating the sustainable operation of a 747 MW CCPP. The thermal efficiency and total power output for the actual, predicted, and simulated models were 27.54% and 667.32 MW, 28.24% and 683.48 MW, and 28.20% and 683.16 MW, respectively. The model with a hidden layer of 10 neurons showed the best results with a Mean Square Error (MSE) of 0.0063237 [14]. Hundi and Shahsavari compared different machine learning models such as Support Vector Regression, ANN, DT, Random Forest, and Gradient Boosting for performance prediction of thermal power plants. Their results demonstrated that all models achieved strong predictive accuracy (R² > 92%), with Random Forest yielding the highest accuracy (~96%) using fewer than half of the approximately 10,000 data points collected from the field [15]. Zhang et al. [16] predicted the electrical load of the CCPP using five different ML methods—Categorical Boosting, Histogram-based Gradient Boosting Regression, Extreme Gradient Boosting Regression, Light Gradient Boosting Machine (GBM), and Support Vector Regression—along with the Hunger Games Search algorithm. As a result of this integration, the combination of Hunger Games Search and Categorical Boosting emerged as the most effective, showing superior performance with an R² value of 0.9735 and a MAE of 2.05525 on the test datasets [16]. Chen et al. [17] proposed a deep learning-based optimization method for the operational control of auxiliary equipment in a combined cycle gas turbine power plant. ML algorithms, including Multilayer Perceptron (MLP), SVM, Gaussian process regression, and linear regression, were used. The MLP model achieved the highest accuracy with the lowest average error of 1.82% [17]. Song et al. [18] aimed to predict the full-load PE of a CCPP using the Super Learner Ensemble technique. This method combines multiple machine learning models to achieve a stronger prediction performance. The study demonstrated that the Ensemble method outperformed individual models, providing more accurate predictions [18].

In this study, the impact of increasing AT and SST due to climate change on Steam Turbine Power Output (ST Power Output) and PE was investigated for a power plant located in a coastal region along the Sea of Marmara in Turkey, where seawater is used as cooling water in the condenser. Regression models were developed to analyze the hourly relationship between V and three input variables: SST, AT, and ST Power Output. Similarly, ST Power Output was correlated with four input variables: V, SST, AT, and RH. Finally, PE was analyzed with five input variables: V, SST, AT, RH, and AP. Despite the abundance of studies in the literature focusing on modeling techniques and analyses for predicting power plant energy output, there is a lack of research specifically addressing the impact of increasing SST on power generation. Therefore, this study distinguishes itself by filling this gap in the literature and contributing to the existing body of knowledge. The analyses utilized a comprehensive real-time dataset comprising 16,006 h of operation at full load from February 2018 to May 2024. The study included extensive error analysis and evaluations of various performance metrics such as MAE, MSE, Mean Absolute Percentage Error (MAPE), RMSE, and Coefficient of Determination (R²).

This research is expected to provide valuable insights for balancing electricity production both in Turkey and globally. Additionally, it may contribute to further research on river-type cooling systems in power plants. Furthermore, this study addresses significant gaps in the existing literature. The key contributions of this paper are summarized as follows:

To determine influencing ST Power Output and PE a 1 °C increase in AT.

To determine the vacuum loss associated with a 1 °C increase in SST.

To examine the impact of vacuum loss on ST Power Output and PE.

To investigate other factors influencing ST Power Output and PE.

To compare predictive models using alternative regression analyses, including regression methods and ANN.

Investigation of the impact of a 1 °C increase in AT and SST on greenhouse gas emissions.

The remainder of this paper is structured as follows: Section 2 provides an overview of the research methodology, explaining the operating principles of CCPPs, the dataset, and research variables. It also describes the regression methods and evaluation metrics used. Section 3 presents the results, analyzing errors and evaluating performance using various metrics. Finally, Section 4 concludes with a summary of the findings.

2. Methodology

2.1. System Description

The CCPP examined in this study has an installed capacity of 950 MW and operates on both Brayton and Rankine cycles. There are two gas turbines and one steam turbine. First, atmospheric air passes through coarse and fine filtration, is compressed by the compressor (1), and then sent to the combustion chamber (2). Meanwhile, natural gas from the Regulating and Metering Station is fed into the combustion chamber, where combustion takes place with the help of the ignitor. The burnt gas is then directed from the combustion chamber to the turbine blades via transition piece (3). After the high-temperature gas releases its energy in the turbine blades, it is directed to the HRSG (4). When the excitation current is applied, a magnetic field is generated in the generator (5), and electricity production begins in the gas turbine. The generator output voltage is increased by a step-up transformer (6), and the generated electricity is directed to the switchyard. High-temperature exhaust gases enable the production of superheated steam at high, intermediate, and low pressures (HP, IP, and LP) and high temperatures in the HRSG. HP and LP steam are sent directly to the steam turbine, and HP steam coming out of the steam turbine (7) is reheated by combining with IP steam in the HRSG and sent back to the steam turbine. The steam that releases its energy in the IP turbine is then sent to the LP turbine and condensed in the condenser (8) using seawater. As with the gas turbine generator, electricity production also begins in the steam turbine generator. The condensate water is sent to the boiler for superheated steam production via the condensate pump (9). The feedwater pump (10) then draws feedwater from the LP drum and sends it to the HP and IP drums. The cycle continues in this manner. The plant’s main cooling water requirement is provided by the seawater pump (11). The schematic of the plant is shown in Figure 1.

In a CCPP, although both AT and SST affect the total power output of the plant, the most significant impact of AT is observed on the Gas Turbine Power Output, while the most significant impact of SST is observed on the Steam Turbine (ST) Power Output. The maximum efficiency of the steam in the turbine is achieved by maintaining a high difference in pressure and temperature between the turbine inlet and outlet. This is achieved by creating a vacuum in the condenser. The quality of the vacuum depends on the type of cooling system used in the plant.

In this study, V, ST Power Output, and PE have been identified as the target variables. Regression analyses were conducted using a dataset of 16,006 h of full-load operation from 2018 to 2024. The flowchart of the research methodology is presented in Figure 2.

Figure 3 shows the histogram graphs of all variables. Hourly average of the ambient variables in the dataset was taken. All variables are defined below.

Ambient Temperature (AT): This input variable ranges from 4.94 °C to 34.60 °C.

Atmospheric Pressure (AP): This input variable ranges from 1003.00 mbar to 1016.00 mbar.

Relative Humidity (RH): This variable ranges from 25.31% to 100.00%.

Sea Surface Temperature (SST): This input variable ranges from 8.36 °C to 25.28 °C.

Vacuum in the Condenser (V): This target variable ranges from 0.025 bar_a to 0.060 bar_a.

Stream Turbine Power Output (ST Power Output): This target variable ranges from 296.10 MW to 321.85 MW.

Full-Load Electric Power Output (PE): This target variable ranges from 834.79 MW to 949.94 MW.

2.2. Regression Methods

In the literature, ML methods have been preferred over thermodynamic methods and mathematical modeling, which require significant computational time and effort to analyze power systems. The ML approach often uses a method called regression to predict values close to reality. ML algorithms are used to estimate a realistic numerical value. Many real-life problems are being solved using ML approaches and algorithms [19].

Regression analysis is a statistical method used to explore the relationship between a dependent variable and one or more independent variables. This method helps in understanding how the independent variables influence the dependent variable. When applied in ML, regression models are powerful tools for predicting outcomes based on one or more predictor variables [20,21].

2.2.1. Multiple Linear Regression (MLR)

Linear regression, introduced in 1894, is a statistical method used to analyze data by examining the strength and nature of the relationships between variables. This technique is commonly applied in modeling scenarios where a dependent variable is predicted based on one or more independent variables, forming either a simple or MLR model [22]. The method assumes a linear relationship between the predictors and the target variable. Its simplicity often makes it ideal for analyzing smaller datasets, as the models are straightforward to interpret. However, linear regression may struggle when dealing with a large number of predictor variables [23].

The basic mathematical formulation of MLR is as follows:

$${\mathrm{Y}}_{\mathrm{i}}={\mathrm{b}}_0+{\mathrm{b}}_1{\mathrm{X}}_1+{\mathrm{b}}_2{\mathrm{X}}_2+\dots +{\mathrm{b}}_{\mathrm{i}}{\mathrm{X}}_{\mathrm{i}}+\mathsf{\epsilon },\mathrm{i}=1,2,3,\dots \dots .,\mathrm{k}$$

(1)

In the equation, Y_i is the observed dependent variable in the MLR, X_i’s are the independent variables, b’s are the regression coefficients, and ε is a constant. Regression coefficient estimates in the MLR model are obtained using the least squares method [24,25].

2.2.2. Trees Regression

Regression trees is a ML method used to develop predictive models from dataset. It is advantageous due to model and interpret complex relationships. The core concept of regression trees is to predict the target variable by making decisions in a tree structure, based on the values of the features in the dataset. This is achieved by repeatedly dividing the data into subsets and applying a simple prediction model to each subset. The process of partitioning is visually represented as a decision tree [26]. The splitting of variables occurs sequentially from top to bottom, meaning that earlier splits are not influenced by subsequent ones [27]. Regression trees are typically used when the dependent variable takes continuous or ordinal discrete values, and prediction errors are generally calculated by the squared difference between the observed and predicted values [26].

2.2.3. Support Vector Machines (SVMs)

Developed by Cortes and Vapnik, SVM is essentially a statistical learning theory and a structural risk minimization method. SVM aims to learn the boundary between classes by separating input values (vectors) into classes [28]. SVM shows strong capability in handling data sets that are not linearly separable or that involve high-dimensional features. Therefore, this model is widely used in statistical classification and regression analysis [29]. SVMs aim to learn an unknown decision function based on a set of N input–output pairs (x_i, y_i). In many real-world scenarios, some prior knowledge about the problem is often available, which suggests that the classifier should account for invariance to factors such as translations and rotations [30].

The fundamental principle behind SVM is to identify an optimal linear hyperplane, framing the search for this hyperplane as an optimization problem through mathematical equations. By leveraging the Mercer kernel’s expansion theorem, SVMs utilize nonlinear mapping to increase dimensionality, thus transforming the original problem into a linear one. This approach effectively addresses complex nonlinear classification and regression issues in the sample space [31].

2.2.4. Kernel Approximation Regression

Kernel Approximation Regression is a term often associated with methods like Kernel Ridge Regression or Kernel Support Vector Regression. These methods use kernel functions to model nonlinear relationships. With kernel approximation, kernel-based algorithms are effectively used in high-dimensional spaces where computation becomes challenging [32]. Kernel methods have become more widely recognized in recent years and are frequently used for pattern recognition and discrimination problems. This sub-module includes functions that approximately compute feature mappings corresponding to specific kernels. Feature functions perform nonlinear transformations of the input, which can form the basis for linear classification or other algorithms [33].

2.2.5. Ensembles of Trees

An ensemble of regression trees is a predictive model that consists of a weighted combination of several regression trees. Typically, using multiple regression trees together enhances the overall predictive accuracy of the model. Ensemble trees provide a unifying approach to handle outputs structured in different ways and build predictive models very efficiently. They can make predictions for various types of structured outputs such as groups of continuous/discrete variables, class hierarchies, and time series [34].

The most popular ensemble of trees methods are Random Forests, GBM, and XGBoost. Random Forests are an extension of Bagging, with the biggest difference being the addition of randomized feature selection [35]. GBMs are defined as a ML technique for regression and classification problems, where an ensemble form of weak prediction models, such as decision trees, is produced. XGBoost, short for Extreme Gradient Boosting, is an ML technique based on gradient boosting and decision tree algorithms, but uses a more advanced algorithm to optimize speed and performance [36].

2.3. Artificial Neural Networks (ANNs)

The ANN is a powerful model used in ML. Neural network techniques are widely used because they do not require many parameters for optimization of results [37,38]. An ANN architecture generally consists of three layers: the input layer, hidden layer, and output layer. The weights connecting the neurons in these layers are structures used for storing information in ANN. The number of neurons, the number of layers, and the types of connections between layers play a significant role in the behavior of an ANN [39]. In a simple ANN architecture, there are only input and output layers, and this is called a single-layer ANN. The weighted sum of the inputs and bias values, combined with an aggregation function, is passed through an activation function (such as sigmoid, hyperbolic tangent, linear function, etc.) to obtain the output. When hidden layers are added between the input and output layers, an MLP architecture is formed. Figure 4 shows the ANN structure. MLPs are classified as Feed Forward Neural Networks (FFNNs) and BPNNs [38,40].

Mathematically, operations in an ANN can be expressed using weight matrices W^(l), bias vectors b^(l), and activation functions σ. For each layer l, the neurons are calculated as follows:

z^(l) = W^(l)a^(l−1) + b^(l)

a^(l) = σ(z^(l))

where, a^(l−1) represents the outputs of the previous layer and a⁽¹⁾ = x represents the initial input.

The hyperbolic tangent activation function has a structure similar to the sigmoid function, but its output values range between −1 and +1. This function is expressed as follows [41]:

$$\mathrm{T}\mathrm{a}\mathrm{n}\mathrm{h}:\mathrm{a}={\displaystyle \frac{{\mathrm{e}}^{\mathrm{z}}-{\mathrm{e}}^{-\mathrm{z}}}{{\mathrm{e}}^{\mathrm{z}}+{\mathrm{e}}^{-\mathrm{z}}}}$$

(2)

2.4. Hybrid Model (LightGBM + DNN)

Traditional single-stage models are effective at learning simple and moderately complex relationships, but they may fall short in capturing more intricate patterns and residual errors. These limitations can be mitigated or even completely eliminated by combining the strengths of different methodologies [42]. The two-stage Ensembles of Trees Residual Learning approach proposed in this study aims to overcome this limitation. In the first stage, a powerful and fast machine learning algorithm, LightGBM, generates the base predictions. Then, a Deep Neural Network (DNN) is employed to learn only the residual errors left by these predictions. In this way, the low-bias advantage of LightGBM is preserved, while the DNN contributes by capturing complex variance, ultimately reducing the total prediction error. This architecture is illustrated in Figure 5. In the initial step, the raw data is read and randomly split into training, validation, and test sets. The hyperparameters of the LightGBM model (number of trees, learning rate, number of leaves, and sampling ratios) are optimized using Optuna to minimize the MAE on the validation set. With the optimal parameters selected, the LightGBM model is retrained on the combined training and validation data to produce the base predictions. Subsequently, the residuals are calculated as the difference between the actual values and the LightGBM predictions. These predictions are added to the input matrix as a new feature column. The extended feature set is then standardized using StandardScaler to have a mean of 0 and a variance of 1. The residual learner, implemented as a DNN, consists of four hidden layers with 128, 64, 32, and 16 neurons, respectively. Each layer utilizes ReLU activation, Batch Normalization, and dropout rates of 30% and 20% in successive layers. The model is trained using the Huber loss function, and the validation performance is monitored using the MAE metric. In the testing phase, final hybrid predictions are generated by summing the LightGBM predictions and the residual corrections produced by the DNN. Model performance is evaluated using RMSE, MSE, R², MAE, and MAPE metrics.

2.5. Prediction Accuracy

2.5.1. Coefficient of Determination (R²)

R² is a statistical metric used to assess the proportion of variance in the dependent variable that is explained by the independent variables in a regression model. It reflects the effectiveness of the independent variables in predicting the variability of the dependent variable. The formula for calculating R² is as follows [43,44]:

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\mathrm{S}\mathrm{S}}_{\mathrm{r}\mathrm{e}\mathrm{s}}}{{\mathrm{S}\mathrm{S}}_{\mathrm{t}\mathrm{o}\mathrm{t}}}}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}}$$

(3)

Here, SS_res and SS_tot are the sum of squared residuals and the total sum of squares, respectively. ${\mathrm{y}}_{\mathrm{i}}$ and $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ represent the actual and predicted values in the dataset, respectively. $\overline{\mathrm{y}}$ is the mean of the actual values, and n is the number of observations.

2.5.2. Mean Absolute Error (MAE)

MAE quantifies the average magnitude of the errors between predicted and actual values. It measures the average absolute deviation without considering the direction of the errors. The MAE is calculated using the following formula [43,45]:

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}⌈{\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}⌉}{\mathrm{N}}}$$

(4)

2.5.3. Mean Absolute Percentage Error (MAPE)

MAPE is a commonly used metric to evaluate the accuracy of a ML model. MAPE measures the average percentage difference between predicted and actual values. It is calculated using the following formula [43,45]:

$$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left|{\displaystyle \frac{{\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}}{{\mathrm{y}}_{\mathrm{i}}}}\right|.100\mathrm{\%}$$

(5)

2.5.4. Mean Square Error (MSE)

MSE is a commonly used metric in regression analysis that calculates the average of the squared differences between predicted and actual values [43,46]. MSE can be computed using the following formula:

$$\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2$$

(6)

2.5.5. Root Mean Square Error (RMSE)

RMSE can be calculated as follows [43,47]:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}\right)}^2}{\mathrm{N}}}}$$

(7)

3. Results and Discussion

This study investigates the impact of rising AT and SST due to global warming on the ST Power Output and the PE of a CCPP using ML techniques based on a series of input variables. For this purpose, various ML regression approaches, including MLR and ANN, were compared as part of a performance evaluation. In the MLR and ANN models, the data were split into training and test sets, with 70.2% used for training and 29.8% for testing in the MLR model, and 69.9% used for training and 30.1% for testing in the ANN model. Based on the comparison results, a hybrid model was proposed to improve performance. Data from 16,006 h of full-load operation at the power plant, covering the period from February 2018 to September 2024, were used for each variable.

3.1. Outliers Detection

A total of 22,253 real system data records were extracted from the Distributed Control System of the CCPP between 1 February 2018 and 5 September 2024. The plant, with an installed net output capacity of 936 MW, has a design value of 950 MW under base load conditions, including internal consumption. Depending on ambient conditions, the base load PE of the CCPP can be as low as 850 MW in the summer, while in the winter, the PE value exceeds 900 MW. During the summer, the base load PE of the CCPP is 850 MW, while in the winter, this value is considered intermediate load. This demonstrates the significant impact of ambient conditions on the target variables, such as V, ST Output, and PE. Under these conditions, outliers were formed. Additionally, it was checked whether the gas turbines were operating under primary or secondary frequency control in the extracted data. Based on all these criteria, outliers were identified using the box-plot outlier method. After removing the outliers, regression analyses were conducted with 16,006 data points. Figure 6 shows the box-plot graphs representing the outliers for the ambient variables.

3.2. Variable Statistics

In this study, each variable has a sample size of 16,006, and

Table 1. Basic statistics of dataset.

	Min	Max	Mean	Std. Deviation
PE	834.79	949.94	888.81	27.78
ST Power Output	296.10	321.85	309.67	4.82
SST	8.36	25.28	17.26	3.27
AT	4.94	34.60	21.01	6.23
RH	25.31	100.00	72.33	12.36
V	0.025	0.060	0.040	0.006
AP	1003.00	1016.00	1009.11	3.28

presents the variable statistics, including the arithmetic mean and standard deviation values.

Among the methods for understanding multicollinearity, the correlation matrix is of great importance.

Table 2. Correlation matrix for V.

	AT	SST	ST Power Output
V	0.366	0.972	−0.739
AT		0.468	−0.791
SST			−0.788

,

Table 3. Correlation matrix for ST Power Output.

	AT	SST	V	RH
ST Power Output	−0.791	−0.788	−0.739	0.262
AT		0.468	0.366	−0.457
SST			0.972	−0.091
V				−0.46

and

Table 4. Correlation matrix for PE.

	AT	SST	AP	V	RH
PE	−0.985	−0.522	0.487	−0.426	0.423
AT		0.468	−0.417	0.366	−0.457
SST			−0.065	0.972	−0.091
AP				−0.009	0.103
V					−0.046

present the correlation matrices showing the relationships between independent variables and the dependent variable. The closer the coefficient between an independent variable and the dependent variable is to 1, the stronger the relationship between the two variables. In Table 2, the value between V and SST is 0.972, indicating a positive effect of SST on V and a very strong linear relationship. In Table 3, the values between ST Power Output and AT, SST, and V are −0.791, −0.788, and −0.739, respectively, indicating that AT, SST, and V have a negative impact on ST Power Output with a relatively high level of correlation. The correlation between RH and ST Power Output is quite weak. In Table 4, the value between PE and AT is −0.985, indicating a negative effect of AT on PE with a very strong linear relationship. The correlations of SST and V with PE are negative and moderate, while the correlation of AP with PE is positive and moderate.

3.3. Effect of Condenser Vacuum on Performance and Analysis of Related Parameters

The distribution of the variables in pairs based on actual and predicted data, using the MLR method, is shown in Figure 7. It can be observed that the relationship between SST and V has the strongest and most linear connection. The performance output of power plants cooled with seawater is higher compared to ACC and tower-type cooled plants. This is due to the slower rise in seawater temperature compared to ambient air temperature. When examining the relationship between ST Power Output and V, it is evident that there is also a strong connection between these two variables. A higher V (lower turbine outlet pressure) positively affects the ST Power Output. However, when examining the relationship between AT and V, the correlation is lower compared to the others. An increase in AT leads to a slight deterioration in V (an increase in turbine outlet pressure). As SST increases, the V deteriorates, and consequently, ST Power Output decreases. As V improves (turbine outlet pressure decreases), ST Power Output increases. Therefore, SST and ST Power Output are the most important input parameters affecting the V output of the CCPP. The effect of the AT input parameter is lower compared to the other variables.

The Multilayer FFNN model consists of an input layer with four neurons, two hidden layers with five neurons, and an output layer with a single neuron responsible for predicting the V value. Based on the results, the number of hidden layers was increased to two in order to enhance performance. The analysis determined that the hidden layer activation function for the Multilayer FFNN method is “hyperbolic tangent”, while the activation function for the output layer is “identity”.

Figure 8 shows the distribution of residuals for V based on the predicted values using the ANN method. Residuals represent the difference between the predicted and actual values. The graph indicates that the errors of the model are largely balanced and randomly distributed, suggesting that there is no significant systematic deviation between the model’s predictions and the actual values.

Figure 9 shows the linear relationship between the values predicted by the ANN model and the actual values. The concentration of data around a linear line indicates that the model makes predictions with high accuracy. The closer the R² value is to 1, the better the performance of the model. The R² value is 0.969, which indicates that the model’s predictions are highly consistent with the actual values.

Table 5. Regression errors with RMSE, MSE, R², MAE, and MAPE performances with three parameters (SST, AT, and ST Power Output) for V.

Methods	Model Type	RMSE	MSE	R2	MAE	MAPE (%)
Multiple Linear Regression	Linear	0.001309	0.000001714	0.9570	0.00096	2.4509
Regression Trees	Fine Tree	0.000977	0.000000954	0.9761	0.00059	1.5178
Support Vector Machines	Cubic SVM	0.001073	0.000001152	0.9711	0.00068	1.7749
Ensembles of Trees	Bagged Trees	0.000843	0.000000711	0.9822	0.00053	1.3589
Artificial Neural Network	Multilayer Perceptron	0.001104	0.000001218	0.9690	0.00074	1.9407
Kernel Approximation	Least Squares Regression Kernel	0.001043	0.000001088	0.9727	0.0007	1.8185
LightGBM + DNN	Hybrid Model	0.000875	0.000000764	0.9806	0.00051	1.322

provides a comparison of RMSE, MSE, R², MAE, and MAPE for the input variables SST, ST Power Output, and AT affecting V, using MLR, ANN, other regression methods, and the hybrid model. It can be observed from the table that the hybrid (LightGBM + DNN) provided the best results.

Figure 10 illustrates the prediction performance of V on both the training and test data using the model obtained with the hybrid method. The points in the prediction graph are positioned very close to the diagonal (45° line), indicating that the model achieves high prediction accuracy on both the training and test data. While the model shows a better fit on the training data, a slight deviation is observed on the test data. However, this deviation remains within an acceptable performance range.

Figure 11 visualizes the prediction performance of SST, AT, and ST Power Output on V using the hybrid model, with distribution plots. The model’s predictions are generally very close to the actual values. It can be seen that as SST increases, the response also increases. Although there is some spread in the response values with AT, the predicted values are still quite accurate. The predictions for ST Power Output generally align with the actual values, but a slightly wider spread can be observed in certain areas.

3.4. Analysis of Related Parameters for ST Power Output and PE

Figure 12 shows the distribution plots of the relationships between four different independent variables (SST, AT, RH, and V) and ST Power Output based on the MLR method. It can be observed that as SST increases, ST Power Output decreases. Similarly, as AT increases, ST Power Output tends to decrease, indicating that higher AT result in a lower steam turbine output power. There is no significant relationship between RH and ST Power Output. As V increases, ST Power Output tends to decrease. The overall result shown by these plots is that SST, AT, and V have a negative effect on ST Power Output, while the RH variable does not show a significant impact.

Figure 13 presents the distribution plots of the relationships between five different independent variables (AT, V, SST, AP, and RH) and PE, based on the MLR method. A linear relationship is observed between AT and PE, where an increase in AT leads to a decrease in PE. The distribution between V and PE shows a scattered dataset, with improvements in V (a decrease in turbine outlet pressure) positively affecting PE. Additionally, as SST increases, PE decreases, while an increase in AP leads to an increase in PE. The relationship between RH and PE is positive but weaker compared to the other variables.

The Multilayer FFNN model consists of a five-neuron input layer and two hidden layers with six neurons for predicting ST Power Output; and a six-neuron input layer and two hidden layers with seven neurons for predicting PE; with a single neuron in the output layer. Figure 14 shows the distribution plots of the actual and predicted ST Power Output and PE obtained with the ANN. The graphs indicate that the predicted values are very close to the actual values.

Figure 15 shows the distribution plots of the effects of Shapley values on ST Power Output and PE using the ANN model. The effect of the SST variable on ST Power Output ranges from approximately −3 MW to 5 MW. The effect of the AT variable on the PE output ranges from approximately −40 MW to 60 MW.

Table 6. Regression errors with RMSE, MSE, R², MAE, and MAPE performances with four parameters (V, SST, AT, and RH) for ST Power Output.

Methods	Model Type	RMSE	MSE	R2	MAE	MAPE (%)
Multiple Linear Regression	Linear	1.8001	3.2403	0.8603	1.3983	0.4526
Regression Trees	Fine Tree	1.7733	3.1445	0.8644	1.2626	0.4091
Support Vector Machines	Cubic SVM	1.6727	2.7978	0.8794	1.2667	0.4104
Ensembles of Trees	Bagged Trees	1.4873	2.2119	0.9046	1.0814	0.3504
Artificial Neural Network	Multilayer Perceptron	1.7172	2.9488	0.8734	1.3352	0.4323
Kernel Approximation	Least Squares Regression Kernel	1.6523	2.7300	0.8823	1.2737	0.4125
LightGBM + DNN	Hybrid Model	1.5032	2.2597	0.898	1.049	0.3399

and

Table 7. Regression errors with RMSE, MSE, R², MAE, and MAPE performances with five parameters (AT, AP V, SST, and RH) for PE.

Methods	Model Type	RMSE	MSE	R2	MAE	MAPE (%)
Multiple Linear Regression	Linear	3.4291	11.7584	0.9848	2.6782	0.30108
Regression Trees	Fine Tree	3.6187	13.0952	0.983	2.7442	0.30943
Support Vector Machines	Cubic SVM	3.2068	10.2834	0.9867	2.5026	0.28162
Ensembles of Trees	Bagged Trees	3.0205	9.1234	0.9882	2.3072	0.26011
Artificial Neural Network	Multilayer Perceptron	3.4365	11.8101	0.9846	2.6731	0.30088
Kernel Approximation	Least Squares Regression Kernel	3.3206	11.0265	0.9857	2.5869	0.29112
LightGBM + DNN	Hybrid Model	2.9005	8.4131	0.9884	2.1942	0.2477

provide a comparison of RMSE, MSE, R², MAE, and MAPE for individual variables for ST Power Output and PE using MLR, ANN, other regression methods, and the hybrid model. It can be observed from all the tables that the hybrid (LightGBM + DNN) method provided the best results.

Figure 16 presents scatter plots comparing the predicted and actual values of ST Power Output and PE in the training and test sets using the hybrid method. For ST Power Output, the predicted points show a linear trend, indicating that the model predicts the actual values quite accurately. For PE predictions, the points are nearly aligned along a straight line, demonstrating that the model makes predictions very close to the actual values. This suggests that the model fits the PE variable very well. Overall, the predicted values of ST Power Output and PE in both the training and test data are quite successful, although small deviations are observed in the test data.

3.5. The Equations Obtained for V, ST Power Output, and PE Using MLR Analysis

The equations representing the V, ST Power Output, and PE of the CCPP based on a set of input variables using MLR analysis are given below:

$$\mathrm{V}=0.06874+0.001836\times \mathrm{S}\mathrm{S}\mathrm{T}-0.000190\times \mathrm{A}\mathrm{T}-0.000181\times \mathrm{S}\mathrm{T}$$

(8)

From Equation (8), it can be seen that a 1 °C increase in SST results in a 0.001836 bar_a increase in turbine exhaust pressure (vacuum degradation). Therefore, its effect on PE is negative.

$$\mathrm{S}\mathrm{T} \mathrm{P}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r} \mathrm{O}\mathrm{u}\mathrm{t}\mathrm{p}\mathrm{u}\mathrm{t}=336.538-350.0\times \mathrm{V}-0.47700\times \mathrm{A}\mathrm{T}-0.0843\times \mathrm{S}\mathrm{S}\mathrm{T}-0.01791\times \mathrm{R}\mathrm{H}$$

(9)

As stated in Equation (9), the primary effect of a 1 °C increase in SST is observed on the vacuum, resulting in a loss of 0.6426 MWh (350 × 0.001836) in ST Power Output. Additionally, other losses arising from the cooling system efficiency (e.g., generator and mechanical losses) have been included in the calculations. Considering the small impact of these additional losses (0.0843 MWh), the total loss in ST Power Output is calculated as 0.7269 MWh.

$$\mathrm{P}\mathrm{E}=72.86-340.3\times \mathrm{V}-4.04681\times \mathrm{A}\mathrm{T}-0.1394\times \mathrm{S}\mathrm{S}\mathrm{T}-0.01779\times \mathrm{R}\mathrm{H}+0.91006\times \mathrm{A}\mathrm{P}$$

(10)

According to Equation (10), the power loss in PE due to the effect of a 1 °C increase in SST on vacuum is calculated as 0.6247 MWh (340.3 × 0.001836). Additionally, when other losses (0.1394 MWh) are included, the total PE power loss reaches 0.7642 MWh. Looking at the coefficient of AT, it can be seen that a 1 °C increase in AT results in a 4.04681 MWh decrease in the CPPP’s PE. Using the PE equation, day-ahead target load predictions have been initiated and yielded successful results in the studied plant. Figure 17, Figure 18 and Figure 19 present a graphical comparison of the actual and predicted results based on the V, ST Power Output, and PE equations according to the MLR method. The graphs were created using 150 data points, and it can be observed that the actual and predicted results are very close to each other.

Table 8. Approach results based on target load according to MAE.

Target Variables	MLR	ANN (MLP)	Best-Performance (Hybrid (LightGBM + DNN))
V	0.0009604	0.0007304	0.00051
ST Power Output	1.3983	1.2839	1.0490
PE	2.6782	2.5780	2.1942

compares the approach values for the target variable according to MLR, ANN, and the best-performance method hybrid. In the comparison made under MAE, it can be observed that the hybrid method provides predictions that are closer to the actual values compared to the other methods. Additionally, it is seen that the ANN performs better than linear regression.

Table 9. Gas Technical Specifications and Greenhouse Emissions.

Gas Components and Ratios
Methane	Ethane	Propan	I-Butane	N-Butane	I-Pentane	N-Pentane	Hexane	N2	CO2	%
95.129683	4.295923	0.256057	0.039296	0.007395	0.007390	0.003412	0.003776	0.259979	0.013644	100.0
Gas Technical Specifications and Greenhouse Emissions
Lower Heating Value, LHV Kcal/sm3	Density, ρ (kg/m3)	Net Calorific Value, NCV (TJ/Gg)		Carbon Content (%)	Consumption (tonne)	Consumption (sm3) (Examined Power Plant Value/ Average Value)		Consumption (Nm3) (Examined Power Plant Value/ Average Value)		EMISSIONS (tCO2e) (Examined Power Plant Value/ Average Value)
8437	0.712027	49.61		0.748023	3186 3409 *	4,475,380 4,788,120 *		4,242,408 4,538,868 *		8733 9344 *

* Average values of CCPPs in Turkey.

presents the characteristics of the natural gas consumed and the emissions resulting from natural gas consumption in the examined CCPP. Standard Cubic Meter (sm³) refers to the volume of gas measured at one atmosphere pressure (1 atm) and 273 K (0 °C) temperature conditions. These conditions are typically referred to as standard conditions. Normal Cubic Meter (Nm³), on the other hand, refers to the volume of gas measured at one atmosphere pressure (1 atm) and 15 °C temperature. This unit is commonly used in daily applications for gas measurement [48]. The CCPP, which operates for an average of 5700 h annually, operates at full load for approximately 5000 h. Equation (10) shows that a 1 °C increase in AT results in a decrease of 4.04681 MWh in PE. This leads to an annual production loss of approximately 20,234 MWh at full load. To produce this energy loss caused by AT, 3,682,588 sm³ of natural gas must be consumed in the same plant. Looking at Equations (8) and (9), a 1 °C increase in SST results in a decrease of 0.7642 MWh in PE, which corresponds to an annual production loss of about 4356 MWh. To produce this loss in the examined CCPP, 792,792 sm³ of natural gas is consumed.

In the examined plant, the total energy loss from a 1 °C increase in AT and SST amounts to 24,590 MWh/year. Table 9 provides the natural gas consumption and greenhouse gas emissions of the examined CCPP and along with the average natural gas consumption and emissions of other CCPPs operating in Turkey.

In the examined plant, approximately 182 sm³ of natural gas is consumed to generate 1 MWh of electricity, resulting in an average emission of 0.355 tCO₂e of greenhouse gases. As a result of global warming, the 24,590 MWh of lost energy, if produced in the examined plant, results in the consumption of approximately 4,475,380 sm³ of natural gas and the emission of 8733 tCO₂e of greenhouse gases.

In natural gas power plants in Turkey, approximately 195 sm³ of natural gas is consumed to generate 1 MWh of electricity, resulting in an average emission of 0.380 tCO₂e of greenhouse gases [49]. If the 24,590 MWh of lost energy were to be produced in another CCPP with average efficiency, then approximately 4,788,120 sm³ of natural gas would need to be consumed, and 9344 tCO₂e of greenhouse gases would be emitted. In this case, an additional 610 tCO₂e of greenhouse gas would have been emitted.

In imported coal and lignite power plants in Turkey, approximately 0.810 tCO₂e and 1.165 tCO₂e of greenhouse gases are emitted, respectively, for every 1 MWh of electricity produced [49]. If the 24,590 MWh of lost energy were to be produced in an imported coal or lignite plant, then 19,918 tCO₂e and 28,647 tCO₂e of greenhouse gases would have been emitted, respectively. In this case, an additional 11,184 tCO₂e and 19,913 tCO₂e of greenhouse gas emissions would have been released, respectively.

When evaluating alternative cooling technologies to prevent energy losses caused by SST, the high installation costs of dry cooling or hybrid systems, combined with the significantly wider annual variation range of AT compared to SST, reduce the efficiency of these systems. Especially during hot seasons, the increase in AT occurs much more rapidly than that of SST. In regions with hot and dry climates, far from the sea, where water scarcity is not an issue, hybrid cooling systems can be used as an alternative to increase efficiency. However, in coastal plants, although the electricity consumption of seawater pumps is significantly higher than that of dry cooling systems, seawater cooling systems stand out due to their positive effects on V.

In addition, since the increase in SST is expected to continue in the coming years, chiller or evaporative cooler systems that increase gas turbine efficiency by lowering the inlet air temperature of the gas turbine, which are mostly used in dry-cooled power plants today, will contribute to reducing the negative impact of SST by becoming widespread in seawater-cooled power plants.

4. Conclusions

In this study, the impact of increasing AT and SST due to climate change on V, ST Power Output, and PE was investigated at a plant located on the coast of the Sea of Marmara, which uses seawater for condenser cooling. The hourly relationships between each target variable and the relevant input variables were analyzed using regression methods, and the model performances were compared. The best results for V, ST Power Output, and PE were obtained using the hybrid model, with MAE values of 0.00051, 1.0490, and 2.1942, respectively.

A 1 °C increase in AT results in a loss of 4.04681 MWh of energy in PE. A 1 °C increase in SST leads to a loss of 0.001836 bar_a in vacuum pressure, which results in a 0.6426 MWh loss in Steam Turbine Power Output. Including other related losses (e.g., generator winding temperature), the total decrease in ST Power Output due to the SST increase is calculated as 0.7269 MWh, and the decrease in PE is 0.7642 MWh. In the examined plant, a 1 °C increase in AT and SST leads to a total energy loss of 24,590 MWh/year. If this lost energy is produced in a CCPP with average efficiency, then an additional 610 tCO₂e of greenhouse gas emissions will occur. If the lost energy is produced from an imported coal plant or a lignite plant, then an additional 11,184 tCO₂e and 19,913 tCO₂e of greenhouse gas emissions will be released, respectively.

The global electricity consumption in 2023 is 29,925 TWh. Of the electricity produced, 35% comes from coal, 22.5% from natural gas, and 9% from nuclear power plants [50]. As of 2024, there are 10,431 oil and gas plants worldwide, with a total capacity of 2136 GW. Of this capacity, 382 GW is from steam turbines [51]. Additionally, there are nuclear power plants with a total capacity of 396 GW and coal-fired plants with a total capacity of 2175 GW in operation. Due to global warming, a 1 °C increase in AT and SST will lead to an annual power loss of approximately 50,000 GWh across power plants. Although the energy lost will vary depending on the production technology (natural gas, coal, lignite), millions of tCO₂e of greenhouse gas emissions will be generated when this energy is reproduced.

However, production losses due to temperature increases (AT and SST), when combined with a reduction in total supply capacity and an increase in demand, can disrupt the supply–demand balance. This situation increases price volatility in wholesale electricity markets and may create temporary supply shortages, especially during peak demand periods. From a national energy policy perspective, it may accelerate investments in reserve capacity, dynamic pricing models, renewable energy integration, and energy storage systems. In addition, temperature-sensitive pricing systems and demand-side management practices may become more widespread.

By integrating renewable energy sources (such as wind and solar) into thermal power systems, hybrid systems can be formed that offer significant environmental and economic advantages. Renewable energy integration can largely offset CO₂ emissions compensated by thermal power plants; however, to fully realize this potential, system flexibility, storage, and transmission infrastructure are equally important. This integration contributes to making today’s energy systems more sustainable, efficient, and reliable.