Non-intrusive load monitoring method with inception structured CNN

Non-intrusive load monitoring (NILM) is an important part of smart grid, which can recognize home electrical appliances. Compared to traditional statistical manners, deep learning can extremely increase the recognition accuracy by more than 10%. However, most NILM methods based on neural networks try to deepen the network to extend the feature extraction capability, which will cause the overfitting and gradient to disappeare in NILM. This paper focuses on solving this problem by optimizing the disaggregation process, and proposes a method based on multiple overlapping sliding windows combined with the inception structure of CNN to disaggregate highly mixed loads of multiple appliances, which can stack each layer disorderly and run each process in parallel, without deepening the depth. Firstly, this work designed a multiple overlap sliding window for NILM to segment the sequence data. Then, the improved inception structure of CNN is used to extract the features, which can provide a rewarding feature extraction capability. After that, the same multiple overlap window is used to smooth the extracted feature of sequence data base on the average filtering. Moreover, this paper makes a comparative analysis of different slide step sizes, which can be concluded that the recognition accuracy is higher when the slide step is shorter. Highly mixed experimental data of multiple appliances is used to test the method. The results highlight the disaggregation performance of the proposed model in the high mixing of multiple electrical load data.