Temperature forecasting plays a critical role in the operation and management of modern energy systems. As one of the most influential meteorological parameters, air temperature directly affects electricity demand patterns, particularly during summer peak periods when cooling requirements dominate total energy use. Variations in temperature can lead to significant fluctuations in electricity consumption, making accurate forecasting essential for optimizing generation schedules, managing reserve margins, and preventing power shortages [10].

The relationship between air temperature and electricity demand is a documented pillar of energy management, with recent data confirming a significant correlation between extreme heat and record-breaking consumption. In Taiwan region, contemporary patterns from 2023 to 2025 show that daily electricity consumption from June to September remains consistently higher than during the rest of the year due to elevated ambient temperatures, with residential and commercial cooling placing substantial pressure on the grid [11]. Consequently, the Percent Operating Reserves (POR) of the grid often decreases as temperatures rise, frequently falling into single digits and triggering supply alerts [5]. These findings emphasize that accurate temperature forecasting is a cornerstone of regional energy security, especially as climate extremes increase the frequency of large-scale outages [6].

The case of Kaohsiung, Taiwan region’s primary industrial hub, further illustrates these stakes. Kaohsiung consistently records high electricity demand driven by heavy industry, dense urbanization, and prolonged high-temperature conditions [12]. Power shortages in such high-demand areas can trigger severe economic losses; for example, the March 2022 nationwide blackout, caused by a failure at the Hsinta Power Plant in Kaohsiung, impacted 5.49 million households and resulted in substantial economic disruption, a modern parallel to the devastating 1999 outage [5]. Comparable events have been observed globally; in the United States, the February 2021 Texas grid collapse (Winter Storm Uri) resulted in estimated economic damages upwards of $130 billion [6].

Global warming brings the importance of temperature effect on energy demand more severely. In 2025, the U.S. experienced 23 separate billion-dollar weather disasters, many involving severe storms and “heat-driven droughts” that threatened regional grid stability across the central and western states. The increasing frequency of these events can be attributed in part to global warming and climate change, which have intensified temperature extremes and made energy systems more sensitive to weather fluctuations. Notably, around 25% of this increase is expected to result directly from climate change, while the remainder stems from socioeconomic development and rising living standards. This growing demand for cooling energy underscores the importance of reliable temperature forecasting in sustainable power system management [13].

Given this context, temperature forecasting during peak electricity demand periods becomes particularly crucial. Reliable temperature forecasts can help power utilities anticipate load fluctuations, plan generation capacity, and prevent blackouts. They also support long-term energy planning and climate adaptation strategies, ensuring that energy infrastructure remains resilient under changing climatic conditions. However, accurately forecasting temperature, especially over long-term horizons, remains challenging due to the nonlinear and dynamic nature of atmospheric processes. These challenges have motivated researchers to explore advanced computational methods, including machine learning and deep learning models, to improve forecast precision and robustness [14].

Forecasting tasks involving sequential or time-dependent data, such as meteorological and energy demand prediction, are highly complex due to the intricate temporal dependencies inherent in the data. Traditional models, including basic Artificial Neural Networks (ANN) and early Recurrent Neural Networks (RNN), frequently encounter challenges when processing long sequences, primarily due to the “vanishing” and “exploding” gradient problems during the training phase. To overcome these limitations, advanced gated architectures such as LSTM and GRU have been developed to effectively capture long-range dependencies.

2.2.1. Long Short-Term Memory (LSTM)

The LSTM network, introduced in [15], is a specialized RNN architecture designed to maintain an internal memory state over extended periods. Unlike standard feedforward networks where inputs are independent, LSTM utilizes a cell state (memory) and a series of “gates” to regulate the flow of information.

Forget Gate: This initial stage determines which information from the previous cell state should be discarded. It utilizes a sigmoid function to evaluate the previous hidden state and the current input, outputting a value between 0 (completely discard) and 1 (completely retain).Input Gate: This gate identifies which new information will be stored in the cell state. A sigmoid layer decides which values to update, while a tanh layer creates a vector of new candidate values; these are multiplied to update the memory.Output Gate: Finally, the output is determined based on the filtered version of the cell state. A sigmoid function decides which parts of the state to output, which is then multiplied by the cell state passed through a tanh function to ensure the final output remains within a standardized range.

While LSTMs are highly effective for complex time-series problems, their sophisticated architecture—consisting of three distinct gates—can lead to increased computational demands and longer training times.

2.2.2. Gated Recurrent Unit (GRU)

As a response to the computational complexity of LSTMs, [16] introduced the Gated Recurrent Unit (GRU), a streamlined variant that offers comparable performance with a simplified structure. The GRU optimizes efficiency by merging the cell state and hidden state and reducing the number of gates to two:

Update Gate: This component combines the functions of the LSTM’s forget and input gates, determining how much of the previous memory should be retained and how much new information should be added.Reset Gate: This gate controls how much of the past information is used to calculate the new candidate state, effectively deciding how to combine the new input with the previous memory.

Because the GRU features fewer parameters, it often achieves faster convergence during training compared to the LSTM, making it an increasingly popular choice for resource-efficient climate modeling. Recent studies have indicated that for multi-year monthly forecasting, the GRU can provide predictive accuracy like that of the more complex LSTM while significantly reducing the number of training epochs required.

In temperature and energy forecasting studies, input variable selection plays a crucial role in determining model accuracy and computational efficiency. Including irrelevant or redundant variables can degrade model performance, while carefully chosen predictors enhance generalization and reduce overfitting. One of the most widely used statistical techniques for identifying relevant input features is correlation analysis, which measures the strength and direction of a linear relationship between variables.

Among various correlation measures, the Pearson Product-Moment Correlation (PPMC) coefficient is commonly employed to quantify the relationship between two continuous variables. The Pearson correlation coefficient ρ is mathematically defined as:

where n is the number of observations; x_i is the value of x (for i-th observation); y_i is the value of y (for i-th observation); µ_x is the mean of x variable; µ_y is the mean of y variable. The r values range between −1.0 and −0.7 shows a strong negative correlation, while a correlation between 0.7 and 1.0 indicates a strong positive correlation. A correlation of 0.0 indicates no relationship at all.

According to statistical conventions follows [17]:

|r| ≥ 0.7: Strong correlation0.5 ≤ |r| < 0.7: Moderate correlation|r| < 0.5: Weak correlation

By applying correlation analysis before model training, redundant and weakly correlated predictors are eliminated. This preprocessing not only reduces computational complexity but also improves model convergence and accuracy. The approach aligns with recommendations from [18] and [19], who emphasized that feature selection based on correlation can significantly improve the performance of forecasting models in energy and environmental applications.

Several measurement indicators have been proposed and employed to assess the prediction accuracy of models. The most frequently used measure is the mean absolute percentage error (MAPE) [18]. A significant advantage of this measure is that it does not depend on the magnitudes of the forecast variables. [19] suggested that MAPE is the most appropriate error measure for evaluating forecasting performance. In this study, the mean absolute percentage error (MAPE) is used to evaluate the forecasting accuracy through cross-sectional analysis.

where A_t is the actual values for period t, F_t the expected value for period t, and n is the number of training samples. The smaller the MAPE, the closer to the actual historical data, the predicted values are, and the more accurate is the forecasting model.

Figure 1 shows the flowchart of LSTM and GRU models which have been used in this investigation. The data set is partitioned to three parts—training data set, validation data set, and testing data set. The training data set was used to determine the optimal forecasting model, the validation data set was used to prevent over-fitting of the various models, and the testing data set was used to investigate the performance of the various models [20]. The optimization scheme of two models is summarized as follows: Step 1 is to use correlation analysis to obtain the best features. Step 2 is to initialize the epoch of the models. Step 3 is to evaluate the fitness of each epoch using MAPE given in Equation (2) for each training and validation data set. Step 4 is to repeat the process, as shown in Figure 1 until the epoch’s number is equal to 1000 with MAPE convergence. When the termination condition is satisfied, the well-trained models and testing data set will be used with the best epoch to investigate the performance of the models.

Figure 1.Flowchart of LSTM and GRU models.

This study focuses on Kaohsiung, Taiwan region’s largest industrial and port city, which consistently records the highest electricity consumption nationwide according to the Taiwan Power Company. Due to its dense industrial concentration and prolonged summer heat exposure, Kaohsiung represents a high temperature-elastic electricity demand environment, making it an ideal case for evaluating forecasting performance under peak-load-sensitive conditions.

The meteorological data used in this study were obtained from the Central Weather Administration (CWA) of Taiwan region, specifically focusing on the Kaohsiung station records from 2010 to 2025. To ensure the reliability of the long-term series, a rigorous completeness check was performed on the hourly records. Although the station identifier transitioned from 467440 to 467441 during the study period because of the updating the equipment from the old to the new, the geographic coordinates and sensor calibration remained consistent, allowing for a seamless integration of the two records. Missing hourly observations accounts for less than 0.5% of the total dataset. Because the study focuses on monthly temperature aggregated from the 10:00 - 17:00 window, which are the working hours for the factories in Taiwan region using the electricity, any days with more than two missing hours within this specific timeframe were excluded from the monthly mean calculation to maintain high data fidelity. This ensures that the final dataset accurately reflects the peak thermal characteristics of the region without being skewed by instrumental downtime or station relocation.

To establish a rigorous data foundation for the forecasting models, the meteorological parameters initially identified in Table 1 underwent a systematic screening process using Pearson correlation analysis. This stage is critical for enhancing model efficiency and predictive accuracy by isolating the most influential drivers of thermal dynamics in Kaohsiung.

To ensure the model adheres to real-world forecasting constraints and to avoid target leakage, all input features were defined as lagged variables. Specifically, for a target temperature at month t, the model strictly utilizes meteorological predictors from the preceding months (t-1 through t-12). As detailed in Table 2, variables demonstrating a strong linear relationship with monthly temperature defined by a Pearson correlation coefficient |r| ≥ 0.7 were prioritized for model integration. These selected predictors, which include station pressure, sea-level pressure, dew-point temperature, and body temperature, serve as the foundational lagged inputs for the LSTM and GRU architectures.

To reduce redundancy and improve model efficiency, Pearson correlation analysis was conducted between monthly temperature and candidate meteorological variables. Variables with absolute correlation coefficients |r| ≥ 0.7 were selected as input predictors.

As detailed in Table 2, the relationships between the potential predictors and the target monthly temperature were quantified, with the strongest correlations highlighted in bold type. These bolded entries indicate variables that meet the selection threshold of a Pearson correlation coefficient |r| ≥ 0.7, signaling a high degree of linear association.

Table 1. Frequently used variables in temperature predictions models.

Implementing this threshold ensures the retention of only the most statistically significant predictors, effectively reducing dimensionality and streamlining the model architecture. By eliminating redundant or weakly correlated variables, this approach improves convergence stability during the estimation process and mitigates the risk of overfitting, thereby enhancing the model’s predictive generalizability.

Table 2.Correlation analysis.

By filtering out weakly correlated variables and strictly utilizing historical lags, the study reduces dimensionality and ensures that the forecasting results are not artificially inflated by contemporaneous data leakage. Finally, all selected variables were normalized using Min-Max scaling prior to model training to ensure numerical stability and accelerate gradient-based optimization:

To preserve temporal integrity and prevent data leakage, the dataset was chronologically divided into three non-overlapping subsets. The training set, spanning the years 2010 through 2022, was utilized to estimate the models’ internal parameters and weights across thirteen years of continuous meteorological observations. Subsequent data from 2023 formed the validation set, which was used exclusively to tune hyperparameters and monitor convergence behavior to prevent overfitting during the training process. Finally, the period of 2024 - 2025 was designated as the testing set, reserved for out-of-sample performance evaluation. In this phase, the models utilized the 2024 data to generate the 2025 temperature forecasts, ensuring that the reported accuracy metrics reflect the models’ ability to predict a truly independent future horizon.

To transition from raw time-series observations to a supervised learning format, a sliding window technique was implemented. This study utilized a look-back window of 12 months, meaning the meteorological features from the previous year were used as input to predict the temperature for the following month, establishing a forecast horizon of t + 1. Given the total dataset span from January 2010 to December 2025, this 12-month sliding window generated a total of 180 overlapping sequences. This sequence count provides a sufficient dataset for deep learning, allowing the LSTM and GRU architectures to effectively learn the recurring seasonal dependencies and long-term thermal dynamics inherent in Taiwan region’s climate. Both models were trained using Backpropagation Through Time (BPTT). The objective of optimization was to minimize Mean Absolute Percentage Error (MAPE). Training continued until validation MAPE stabilized, preventing overfitting. The parameter settings of LSTM and GRU are shown in Table 3.

Table 3.Parameter setting of the optimization searching algorithms.

To ensure a rigorous and controlled comparison, both the LSTM and GRU architectures were evaluated using an identical hyperparameter search space. This included a shared configuration for the Adam optimizer, a consistent batch size of 32, and a standardized unit count of 50 per layer. While the LSTM model was configured with three recurrent layers, the GRU architecture utilized two. This structural variation was determined through preliminary hyperparameter tuning on the validation set; it was observed that a third GRU layer introduced unnecessary architectural complexity that led to immediate overfitting and increased computational latency without improving predictive accuracy. Therefore, the chosen configurations represent the optimal depth for each specific model type under the same experimental constraints, ensuring that the performance comparison reflects the best-tuned version of each architecture.

Based on the empirical observations during the training phase, the two recurrent neural network architectures exhibited distinct convergence behaviors. The Long Short-Term Memory (LSTM) network reached convergence at approximately 800 epochs, whereas the Gated Recurrent Unit (GRU) network converged more rapidly, requiring only about 400 epochs.

To ensure optimal forecasting performance and mitigate the risk of overfitting, the final weights for both models were not necessarily taken from the final epoch. Instead, the models were strictly selected based on the epoch that yielded the minimum Mean Absolute Percentage Error (MAPE) on the validation set.