Deep residual networks with convolutional feature extraction for short-term load forecasting

Conventional deep learning models struggle with balancing feature extraction and long-term temporal representation in Short-Term Load Forecasting (STLF). This study proposes a Convolutional Neural Network-Embedded Deep Residual Network (CNN-Embedded DRN) designed to enhance early-stage feature extraction and generalization capability across diverse climatic conditions. The objectives of this study are to integrate Convolutional Neural Network (CNN)-based local feature extraction into the DRN framework for capturing fine-grained temporal and spatial load patterns, to employ residual learning for mitigating gradient degradation and improving network stability, to evaluate the model's predictive performance against baseline and ablation models across two datasets representing temperate (ISO-NE) and tropical (Malaysia) climates, and to validate its statistical significance and seasonal robustness through bootstrap analysis and multi-seasonal evaluation. The results demonstrate that the proposed CNN-Embedded DRN consistently outperforms all comparative models, achieving the lowest Mean Absolute Percentage Error (MAPE) values of 1.5303% and 5.0566% on the ISO-NE and Malaysia datasets, respectively. The inclusion of residual network (ResNet) and CNN-Embedded ResNet as ablation experiments confirms that CNN-based local feature extraction effectively complements residual learning, while bootstrap analysis verifies that the observed improvements are statistically significant. The proposed model provides a reliable and generalizable framework for STLF, offering improved accuracy, robustness, and adaptability under varying climatic and demand conditions. Future research will focus on extending this framework toward multi-regional and multi-scale forecasting, incorporating attention mechanisms for enhanced long-term dependency modeling, and exploring adaptive hybrid residual architectures for real-time energy management applications.