A Novel Short‐Term Residential Load Forecasting Model Integrating Holiday and Periodic Patterns

ABSTRACT

To address the current challenges in short‐term residential load forecasting, this study proposes a novel model named the Learnable Recurrent Cycle Fusion‐Holiday Encoder Patch Lightweight Adaptive Network (LRCF‐HEPatchLAN). By integrating holiday information and periodic patterns, this model achieves highly accurate load forecasting. First, considering the data complexity, noise interference, and anomalous fluctuations inherent in residential loads, an improved Patch Lightweight Adaptive Network (PatchLAN) model is developed based on the PatchTST framework. It utilizes the robust noise processing and adaptive feature extraction capabilities of A Lightweight Adaptive Network (LANet) to enhance forecasting performance. Second, to capture periodic load characteristics, a Learnable Recurrent Cycle Fusion (LRCF) technique is designed, which extracts cyclical trends and integrates periodic patterns through trainable recurrent cycles. Finally, a smoothed holiday encoding mechanism is introduced to embed holiday information into the input, optimizing adaptability to special dates. experiments on three load datasets from diverse regions in China demonstrate that the proposed model outperforms others in short‐term forecasting tasks. Specifically, it reduces the Mean Absolute Error (MAE) by 15.69%–32.52% and the Root Mean Square Error (RMSE) by 11.25%–28.74%. Furthermore, experimental analysis verifies the rationality of the model design, demonstrating that the improved PatchLAN effectively handles noise and enhances model robustness, the LRCF strengthens the capture of periodic trends, and the holiday encoding mechanism improves adaptability to load variations on special dates.