Deep residual networks for short-term load forecasting: an empirical study on the impact of network depth

The impact of network depth on the performance of deep residual networks (DRNs) for short-term load forecasting (STLF) remains insufficiently understood. This study conducts a systematic empirical investigation of network depth within DRN-based frameworks, with a particular focus on its relationship with forecasting performance, model behavior, and dataset characteristics. Two representative models, namely the original DRN and the Principal Component Analysis–Deep Residual Network (PCA–DRN), are evaluated under a wide range of depth configurations using two real-world datasets with different load and meteorological characteristics: ISO-NE and MyPJ. The results suggest that the influence of network depth on forecasting performance is inherently nonlinear and highly dataset-dependent. For the ISO-NE dataset, medium-depth configurations achieve favorable performance by effectively modeling more pronounced seasonal variability and long-term temporal patterns. In contrast, relatively shallow configurations generally exhibit improved performance on the MyPJ dataset under the current experimental setting, suggesting that deeper architectures may introduce additional model complexity without consistent performance gains. Furthermore, within the MyPJ dataset containing multiple meteorological variables, the incorporation of PCA improves feature representation, enhances model robustness, and is associated with reduced sensitivity to depth variations under the current experimental setting. Comparative evaluations against mainstream deep learning (DL) models indicate that DRN-based frameworks achieve competitive forecasting performance with relatively efficient structural designs. In addition, bootstrap-based statistical analysis suggests that the observed performance differences remain distinguishable under sample-level variability within the current experimental setting. These findings suggest that appropriate network depth selection should be considered together with dataset characteristics and feature representation, providing practical insights for the design of efficient and robust STLF models.