Non-intrusive Load Identification Based on CNN-DBN-BiLSTM

The rapid development of smart grid has promoted the application of non-intrusive load monitoring (NILM) technology in energy management. Aiming at the problems of insufficient feature expression and information loss in existing algorithms, insufficient generalization ability due to scarcity of labeled data, weak model anti-noise ability and low load recognition accuracy, this study introduces a deep learning framework integrating CNN, DBN, and BiLSTM for identifying appliance loads in a non-intrusive manner. Firstly, CNN is used to capture the local spatial characteristics of load transient power to enhance the model’s perception of equipment operation mode; secondly, DBN is introduced to extract global high-order abstract features from unlabeled data through unsupervised pre-training mechanism; finally, BiLSTM is combined to capture the long-term operation characteristics of multi-stage electrical appliances, and the coordinated optimization of local details and global time series features is achieved. Experiments show that the recognition accuracy of 10 types of electrical appliances on the UK-DALE dataset is $95.70 \%$, which is $8.41 \%$ higher than that of traditional CNN, BiLSTM and CNN-BiLSTM models. The loss value converges to $\mathbf{0. 0 0 0 6 5}$, which verifies the high efficiency and strong generalization ability of the model.