Abstract
Short-Term Load Forecasting (STLF) is essential for ensuring the stability and operational efficiency of modern power systems. Deep Residual Networks (DRNs) have recently demonstrated promising performance in this domain, enabling more effective training of deeper architectures. However, existing activation functions such as the Scaled Exponential Linear Unit (SELU) rely on fixed parameters and strict initialization, which may limit their adaptability to seasonally varying load–weather conditions. This study introduces a modified activation function, the Parametric Scaled Exponential Linear Unit (PSELU), which incorporates a small tunable parameter γ in the negative region, extending the formulation of SELU while preserving its self-normalizing characteristics. Experiments conducted within the DRN framework on two benchmark datasets—ISO-NE (temperate climate) and Malaysia (tropical climate)—demonstrate that the DRN model employing the proposed PSELU (γ = 0.02) achieves modest yet consistent improvements in forecasting accuracy compared with the DRN model using SELU. Specifically, the Mean Absolute Percentage Error (MAPE) decreased from 1.718% to 1.662% for ISO-NE and from 5.251% to 5.012% for Malaysia. Although the improvements are moderate, they were statistically validated through 10,000-iteration Bootstrap resampling at the 95% confidence level. These results suggest that the limited parameterization of SELU enhances consistency and adaptability in forecasting performance across different climatic and seasonal conditions. Future work will expand the evaluation to a wider range of datasets and model architectures to further examine the generalizability and practical applicability of PSELU in diverse forecasting contexts.