An integrated method for electric load forecasting in large public buildings based on load patterns clustering and multi-scale temporal feature selection

Existing power load forecasting models for large public buildings generally lack targeted mechanisms for pattern recognition and feature selection when dealing with complex and variable load patterns, resulting in insufficient prediction stability and generalization ability. To address this issue, this paper proposes a novel forecasting method that integrates Multi-Scale Dynamic Time Warping (MS-DTW), Feature Importance Evaluation (FIE), Improved Temporal Convolutional Network (ITCN), and Transformer, combining load pattern clustering, feature importance assessment, and an enhanced temporal modeling structure. Experimental results show that: (1) the MS-DTW-based load pattern clustering effectively supports segmented modeling and demonstrates strong adaptability in three representative scenarios selected in this study, which are significantly influenced by holidays and seasonal cooling loads; (2) feature importance analysis reveals that load, hour of the day, and month are the most critical factors affecting prediction accuracy, while seasonal features contribute relatively little. Across the three typical building types and forecasting horizons of 6, 24, and 48 h, the proposed MS-DTW-FIE-ITCN-Transformer model consistently outperforms both baseline and ablation models. In the 6-h forecasting task, compared with the best-performing baseline model TCN-Transformer, RMSE is reduced by 39.2% in Hebei hospital, 47.1% in Shaanxi university campus, and 44.6% in Zhejiang commercial complex. Moreover, the R2 values in all scenarios exceed 0.9, validating the effectiveness of the proposed method.