Mini-batch size sensitivity in deep residual networks for short-term load forecasting: an empirical study.

Mini-batch size is a fundamental yet often empirically selected training configuration in deep learning (DL)–based short-term load forecasting (STLF). Despite the widespread adoption of deep residual networks (DRNs) for STLF, the sensitivity of forecasting performance to mini-batch size has received limited systematic investigation. This study presents a comprehensive empirical analysis of mini-batch size effects in DRN-based STLF, using the original DRN and the Principal Component Analysis–Deep Residual Network (PCA-DRN) as representative frameworks. Experiments are conducted on real-world electricity load datasets representing different climatic conditions under a unified experimental setting, where network architecture, loss function, optimizer, and training epochs are strictly controlled and only the mini-batch size is varied. The results demonstrate that both DRN and PCA-DRN exhibit pronounced sensitivity to mini-batch size, with medium-scale batches achieving superior forecasting performance. In particular, a mini-batch size of 64 consistently yields the best results, while a batch size of 32 provides a robust alternative. Comparative evaluations with representative DL baseline models further highlight the effectiveness of residual-based forecasting. Bootstrap-based statistical significance testing confirms that the observed performance improvements are reliable and reproducible. Overall, this study provides practical guidance on mini-batch size selection for residual-network-based STLF and offers new insights into training configuration effects in DL-driven STLF across different climatic electricity systems.