Deep Neural Network in Sequence to Short Sequence Form for Non-intrusive Load Monitoring

Recently, the deep learning method of sequence to sequence (seq2seq) form has achieved excellent results in non-intrusive load monitoring(NILM). However, due to the particularity of NILM task, which requires continuous output of appliance power, when using this form, there will be a situation that only part of the operation time in appliances will be included in the input sequence of meter power, resulting in a decline in the effect of energy disaggregation. In this paper, a convolutional neural network in the form of sequence to short sequence (seq2sseq) is proposed. The seq2sseq network only outputs a short appliance power sequence corresponding to the middle position of the input sequence. As more input information outside the time period of the output sequence can be seen, the problem of inaccurate estimation in the sequence edge can be solved. The experiments show that the seq2sseq-based neural network can achieve better results than the seq2seq-based one while ensure computational efficiency. It is a more suitable form of neural network for NILM task.