Understanding the influence of training schedules on short-term load forecasting via deep residual networks: An empirical study

Modern electrical systems rely on short-term load forecasting (STLF) for reliable and efficient operation. Although deep residual networks (DRNs) have achieved strong forecasting performance in STLF, most existing studies mainly emphasize architectural modifications, while the influence of training strategies has received relatively limited attention. In particular, the effect of training schedule design, especially epoch allocation, remains insufficiently studied. This work examines the influence of epoch allocation on DRN-based STLF under a unified multi-stage snapshot ensemble framework. By keeping the model architecture and optimization settings unchanged, the analysis focuses specifically on the effect of different epoch allocation strategies on convergence behavior and generalization performance. Experiments are conducted using two real-world datasets, namely the New England Independent System Operator (ISO-NE) dataset and the Malaysia Petaling Jaya (MyPJ) dataset, which exhibit different temporal characteristics. The experimental results show that the optimal training schedule depends strongly on dataset properties. The ISO-NE dataset benefits from longer training durations and later snapshot extraction stages, whereas the MyPJ dataset achieves better forecasting performance at earlier training stages because of stronger short-term variability. In addition, principal component analysis (PCA)-based feature representation improves forecasting stability and reduces forecasting error by alleviating redundancy among meteorological variables. Statistical analysis further confirms that these improvements are statistically significant. Overall, the results demonstrate that suitable training schedule design can improve DRN-based STLF performance without increasing model complexity.

Text 128137.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) Official URL or Download Paper: https://www.nature.com/articles/s41598-026-54556-3

Actions (login required)

View Item