关键词: 非侵入式负荷监测, 负荷识别, 非电气量特征, 特征融合, Timesnet模型, 注意力机制

Abstract: To address the issue that existing non-intrusive load monitoring (NILM) models over-rely on the inherent power characteristics of electrical appliances, which restricts the improvement of monitoring performance, this study introduces the impacts of environmental factors (e.g., temperature, humidity, and wind speed) on the usage frequency and operational modes of electrical devices. A novel TimesNet-based multi-source feature fusion method for non-intrusive load monitoring is proposed. Specifically, the TimesNet model is leveraged to accommodate the multi-periodic characteristics of load power sequences. Meanwhile, an Inception network is adopted to extract the contextual correlations of aggregate load data, thereby enhancing the capability of capturing multi-scale periodic features of power sequences and deriving power characteristics that are more consistent with the operational patterns of electrical appliances. Secondly, to tackle the temporal pattern heterogeneity and dimensional differences between power sequences and environmental sequences, a multi-head self-attention mechanism is employed to achieve deep fusion of power features and environmental features (including temperature, humidity, and wind speed), which effectively excavates the potential correlations among multi-source features. The proposed model is validated against six representative models on the AMPds2 and UK-DALE datasets. Experimental results demonstrate that the proposed model outperforms the comparative models in terms of mean absolute error (MAE) and root mean square error (RMSE). Particularly, it achieves a remarkable improvement in identification accuracy for environment-sensitive or strongly periodic appliances such as heat pumps and refrigerators. These results confirm that the proposed model can effectively enhance the overall load identification accuracy, significantly mitigate the error caused by over-reliance on appliance power characteristics, and exhibits excellent load monitoring performance.

Key words: Non-invasive load monitoring, Load identification, Non-electrical quantity characteristics, Feature fusion, Timesnet model, Attention mechanism