MSDCANet: A multi-scale dual-channel convolutional attention network for non-intrusive load disaggregation with enhanced feature extraction

Abstract

Non-Intrusive Load Monitoring (NILM) is a core smart grid technology that identifies the operating states of multiple appliances from a single access point, making it highly valuable in energy management and conservation. Although deep learning-based NILM methods have significantly improved feature extraction, the performance bottleneck has shifted beyond mere extraction accuracy as neural networks advance. Most models rely on single-scale features, overlooking multi-scale variations in power data caused by operational mode transitions. Appliance operation inherently exhibits multi-scale characteristics, with features at different scales jointly determining behavior; single-scale extraction may lead to overfitting and hinder differentiation of appliances with similar patterns. To address these limitations, we propose a Multi-Scale Dual-Channel Convolutional Attention Network (MSDCANet) for NILM tasks. MSDCANet integrates multi-scale feature extraction, adaptive normalization, and a multi-scale attention mechanism to extract and fuse features at various scales, enhancing disaggregation accuracy and model generalization. We evaluate it under origin-household and cross-household paradigms on the UK-DALE and REDD datasets. Experimental results demonstrate that MSDCANet outperforms state-of-the-art models in MAE, SAE, and F1 metrics for several high-energy-consuming appliances, confirming its applicability in complex usage scenarios and underscoring the importance of multi-scale techniques in NILM. The source code for our work is available at https://github.com/linfengYang/MSDCANet.