การเปรียบเทียบประสิทธิภาพการพยากรณ์อนุกรมเวลาปริมาณการใช้กระแสไฟฟ้าสูงสุดรายเดือนด้วยแบบจำลองปัญญาประดิษฐ์

Abstract

The purposes of this research were to develop and compare the performance of monthly peak electricity load time series forecasting models using five artificial intelligence models, namely: (1) Recurrent Neural Network (RNN), (2) Convolutional Neural Network (CNN), (3) Long Short-Term Memory (LSTM), (4) Hybrid RNN-LSTM, and (5) Hybrid CNN-LSTM. The data used in this research consisted of the monthly peak electricity load time series data of Thailand from 1 January 1986 to 31 December 2025. The dataset was divided into two subsets, including 408 records for model development and training, and 72 records for model testing. The results indicated that the Hybrid RNN-LSTM model achieved the highest overall forecasting performance. During the training phase with the training dataset, the model yielded the lowest Mean Absolute Error (MAE) of 423.89 MW, the lowest Root Mean Squared Error (RMSE) of 588.92 MW, and the lowest Mean Absolute Percentage Error (MAPE) of 2.47%, while achieving the highest Coefficient of Determination (R²) of 0.9932. During the testing phase with the testing dataset, the model also produced the lowest MAE of 988.47 MW, the lowest RMSE of 1,243.24 MW, and the lowest MAPE of 3.29%, while achieving the highest R² value of 0.7411. In addition, the forecasting accuracy estimation showed that the Hybrid RNN-LSTM model achieved the lowest Mean Magnitude of Relative Error (MMRE) of 4.36%. The findings indicated that the Hybrid RNN-LSTM model could effectively learn the patterns of time series data and provided better forecasting performance for monthly peak electricity load time series than the other models used in this study.