Adaptive user clustering enhanced BiLSTM-attention for short-term load forecasting in smart distribution networks

Multi-node load forecasting in distribution networks faces user heterogeneity and computational efficiency bottleneck challenges, with existing static clustering and global models struggling to balance forecasting accuracy with real-time requirements. This work proposes an adaptive clustering enhanced two-layer BiLSTM-Attention collaborative forecasting framework based on DTW-K-Medoids, employing dynamic time warping similarity metric for robust load pattern identification, implementing a 4-week sliding window adaptive update mechanism to address seasonal load evolution, and constructing a cluster-wise training strategy enabling models to focus on temporal modeling of homogeneous user groups. Experiments on the ISO New England dataset, with cross-dataset validation on GEFCom2014, confirm that this method significantly outperforms baseline methods in both prediction accuracy and computational efficiency, with ablation experiments quantitatively revealing independent contributions of each component and scalability analysis confirming computational advantages in large-scale distribution networks. This research provides a solution balancing accuracy and efficiency for day-ahead load forecasting in smart distribution networks, with the cluster-wise training strategy effectively addressing real-time prediction demands for large-scale nodes through reduced computational complexity, offering data support for grid optimal scheduling and demand-side management.