Non-intrusive load disaggregation based on composite deep long short-term memory network

Abstract

Non-invasive load monitoring (NILM) is a vital step to realize the smart grid. Although the existing various NILM algorithms have made significant progress in energy consumption feedback, there are still some problems need to further addressed, such as the exponential growth of state space with the increase of the number of multi-state devices, which leads to the dimension disaster; and it is difficult to capture the power fluctuation information effectively because of the neglect of time-dependency problem load disaggregation; traditional disaggregation involves a process of one sequence to one sequence optimization, which is inefficient. In our study, a composite deep LSTM is proposed for load disaggregation. The proposed algorithm considers the process of load disaggregation as a signal separation process and establishes regression learning from a single sequence to multiple sequences to avoid dimension disaster. In addition, an encoder-separation-decoder structure is introduced for load disaggregation. Encoder completes the effective encoding of the mains power and differential power information, the time-dependency of the encoding process implemented by a deep LSTM, separation realizes the disaggregation process by separating the encoded information, and decoder decode the separated signal into the sequences of corresponding electrical appliances. Compared with the one sequence to one sequence disaggregation method, the proposed method simplified disaggregation complexity and improves the efficiency of disaggregation. The experimental results on WikiEnergy and REDD datasets show that the proposed method can reduce the disaggregation error and improve the comprehensive performance of event detection. Besides, our study can provide conditions for the realization of the bidirectional interaction of the smart grid and the improvement of the smart grid scheduling.