Optimizing Deep Residual Networks for Short-Term Load Forecasting With Multidimensional Weather Data and Principal Component Analysis

Short-term load forecasting (STLF) is essential for power system operations, supporting efficient grid management and resource planning. Deep Residual Networks (DRNs) have emerged as a promising architecture for STLF, offering a balanced solution by combining training stability, deep feature extraction, and reduced gradient degradation compared to Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers. This study focuses on optimizing DRN for STLF by evaluating different combinations of multidimensional weather variables and applying Principal Component Analysis (PCA) to address feature complexity. Using the Malaysia dataset, which includes historical load, time, and weather variables such as temperature, rainfall, and wind speed, the impact of different variable combinations on forecasting precision is evaluated. Experimental results show that the DRN model outperforms baseline models including CNNs, RNN-based models, and Transformers, achieving a Mean Absolute Percentage Error (MAPE) of 0.052514 and a coefficient of determination (R2) of 0.927993. Building upon this, the proposed PCA-DRN further improves forecasting performance, achieving a MAPE of 0.049994 and an R2 of 0.934473, representing a 4.80% reduction in MAPE and a 0.65% increase in R2 compared to the original DRN. These findings emphasize the importance of feature selection and dimensionality reduction in optimizing STLF models, particularly for tropical regions with relatively stable weather patterns.