基于层次化特征解耦迁移的增量用户负荷精细画像方法OA

To address the performance degradation in incremental user load profiling caused by data scarcity,an incremental user load fine-grained profiling method based on hierarchical feature decoupling and transfer is proposed.This method treats data-abundant existing users as the source domain and provides profiling support for incremental users through hierarchical feature disentanglement and cross-domain alignment.First,a shared encoder integrating multi-scale convolution,Transformer and bidirectional long short-term memory(BiLSTM)networks is pre-trained to extract multi-scale temporal features from load sequences.Second,a hierarchical feature disentanglement module decomposes temporal features into domain-invariant,cluster-shared and domain-specific components at the global,group and individual levels.The decomposed features characterize common consumption patterns,group-level load modes and individual consumption behaviors respectively.Then,a dynamic clustering mechanism is introduced to match incremental users with similar existing user groups,and cross-domain feature alignment is achieved by combining contrastive learning and adversarial training.Finally,the encoder parameters are frozen and the profiling layer is fine-tuned using a small amount of target domain data.Experimental results on real load data from a city in southern China show that the proposed method outperforms conventional transfer learning methods in incremental user load forecasting,demonstrating its effectiveness for load pattern modeling and transfer adaptation.