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:

Current non-intrusive load monitoring （NILM） models rely excessively on appliance power features， resulting in limited monitoring accuracy under complex environmental conditions. To address this problem， a multi-source feature fusion NILM model based on TimesNet and the attention mechanism， namely TAMF， was proposed. By introducing non-electrical environmental factors， including temperature， humidity， and wind speed， TAMF mitigated errors caused by dependence on single power features. TimesNet was used to adapt to the inherent multi-periodicity of load power sequences. A one-dimensional power sequence was transformed into a two-dimensional tensor through fast Fourier transform， and contextual relationships and multi-scale periodic features of aggregate load data were extracted with the multi-branch convolutional structure of the Inception network， yielding power features more consistent with actual appliance operation characteristics. Considering that environmental features showed more obvious local short-term fluctuation patterns than power features， a one-dimensional convolutional neural network with fewer parameters and lower computational complexity was used to extract local temporal features from environmental sequences， including temperature， humidity， and wind speed. Subsequently， to address heterogeneity of power and environmental sequences in temporal patterns and dimensions， a multi-head self-attention mechanism was introduced. Parallel multi-head computation enabled deep fusion of power and environmental features， extracted potential associations and interactions between two types of multi-source heterogeneous features， and improved load identification accuracy of target appliances. The AMPds2 dataset from Vancouver， Canada， was selected as the main experimental dataset， and the UK-DALE dataset from London， United Kingdom， was used to verify cross-dataset generalization ability. TAMF was compared with two traditional methods， combinatorial optimization （CO） and factorial hidden Markov model （FHMM）， as well as four deep learning models， sequence-to-point （Seq2point）， gated recurrent unit （GRU）， bidirectional encoder representations from Transformers （BERT）， and Transformer. The results showed that mean absolute error and root mean square error of TAMF on both datasets were lower than those of all six comparative models. On the AMPds2 dataset， mean absolute error and root mean square error decreased by 17.27% and 4.73%， respectively， compared with the best comparative model. On the UK-DALE dataset， they decreased by 14.02% and 6.51%， respectively. The improvement in identification accuracy was especially significant for environmentally sensitive appliances such as heat pumps and refrigerators. In addition， ablation experiments further confirmed that the synergistic effects of the feature fusion network and environmental feature extraction network were key to improving overall model performance， and both networks complemented the power feature extraction network. In summary， introducing environmental factors and realizing deep fusion of multi-source features can effectively reduce identification errors caused by dependence on single electrical features， improve overall load identification accuracy， and verify that TAMF maintains excellent load monitoring and generalization ability in complex scenarios with differences in data distributions and feature dimensions. This model provides a practical new approach for improving non-intrusive load monitoring performance， supporting urban residential electricity-consumption behavior analysis， and enabling low-cost customer-side load perception in smart grids.

Key words: non-intrusive load monitoring, load identification, non-electrical features, feature fusion, TimesNet model, attention mechanism