A Training-Enhanced NILMFormer for Non-Intrusive Load Disaggregation Under Diverse Appliance Load Structures

Abstract

Existing sliding-window deep-learning models for non-intrusive load disaggregation typically use a unified training objective and fixed training configurations, which limits their ability to accommodate structural differences among appliance loads. To address this limitation, this article proposes a training-enhanced Non-Intrusive Load Monitoring Transformer (NILMFormer) for load disaggregation under diverse appliance load structures. Guided by the principle of matching the training process to appliance load structures, the proposed method establishes a Load-Structure-Matched Coordinated Enhancement Framework (LSM-CEF). First, an Adaptive Scale-Routed Local Expert Adapter (ASR-LEA) is incorporated into the feature representation stage to strengthen local multi-scale feature extraction and improve the characterization of critical appliance patterns. Second, a Huber robust regression loss is introduced to reduce the influence of occasional spikes and local abnormal deviations in real-world smart meter data during parameter optimization. Third, a mapping mechanism from load profiles to training configurations is developed. Based on differences in point-level ON ratio, window-level sparsity, and event persistence, the mechanism adaptively selects suitable training enhancement strategies for different appliances. Finally, boundary-enhanced point-wise weighted regression and key-window resampling mechanisms are designed to improve the learning of critical power variation processes at both the time-point and window levels. Experiments are conducted on the public UK-DALE and REFIT datasets. Compared with the baseline NILMFormer model (NILMFormerBase), LSM-CEF reduces the average MAE by 47.02% and improves the average MR by 85.28% on UK-DALE, while reducing the average MAE by 33.01% and improving the average MR by 64.14% on REFIT. The results demonstrate that the proposed method effectively improves power reconstruction accuracy and overall energy matching capability across diverse appliance load structures.