Short-term power load forecasting using deep learning method with multi-day weather cumulative effects

Accurate and reliable short-term power load forecasting is crucial for the secure and economic operation of modern power grids. Existing models often overlook the cumulative physiological effects of meteorological factors, leading to significant degradation in predictive accuracy during persistent extreme weather events such as heatwaves. To address this limitation, this study proposes a novel forecasting framework driven by a synergy of physical mechanisms and deep learning. A multi-day weather cumulative effect index ( E-index ) is defined to explicitly quantify the delayed feedback of meteorological features over time based on human thermal adaptation priors. A hybrid deep learning architecture, VMD-KAN-LSTM-Attention, is developed. Variational Mode Decomposition (VMD) mitigates high-frequency noise, while the integration of Kolmogorov-Arnold Networks (KAN), Long Short-Term Memory (LSTM) networks, and an attention mechanism efficiently captures highly nonlinear mappings and dynamic temporal dependencies. Systematic evaluations reveal that the proposed E-index acts as a universally effective physical prior, consistently enhancing the predictive accuracy across diverse algorithmic architectures, including XGBoost and CNN-BiLSTM. The integration of the E-index yields the most significant performance improvements during peak summer periods by explicitly capturing multi-day thermal accumulation. Consequently, the VMD-KAN-LSTM-Attention framework empowered by the E-index demonstrates superior robustness and significantly outperforms state-of-the-art baselines across all evaluation metrics. This study provides a new theoretical methodology and a reliable engineering tool for meteorologically sensitive load forecasting in modern power systems.