SingleNet: An end-to-end single network for non-intrusive load monitoring of multiple appliances

An end-to-end convolution neural network (CNN) model SingleNet is proposed for non-intrusive load monitoring (NILM) in this paper. A SingleNet can identify several typical types of household appliances, and give the probability, start time, stop time and the average power of the appliances. The network can receive data input with variable-length by using multi-scale pooling technology. A new loss function is proposed by combining classification loss, positioning loss, and power loss with the help of cross entropy, smooth L1, and Euclidean distance respectively. The new loss function can better converge the training process and significantly shorten the training time. Finally, The SingleNet achieves an end-to-end recognition effect by inputting variable-length total-power sequence data, and outputting the category, start time, stop time and average power of household appliances.