Wavelet Decomposition-bi-directional Long-short Term Memory Neural Network with Prob-sparse Attention for Non-intrusive Load Monitoring

Energy disaggregation or non-intrusive load monitoring (NILM) aims to identify the operating state of every single appliance from given aggregate energy. Many methods are only based on time information, which have difficulties in identifying multi-state processes of an appliance as important information may be lost when the sequence is too long. Considering frequency domain features offer additional information and have the advantage of data compression, this paper proposes to make full use of time-frequency information. First, a time series of aggregated power is reshaped into a sequence of 2D images, and then wavelet decomposition (WD) is employed to extract time-frequency information in horizontal, vertical, and diagonal directions. Afterwards, information is processed by a bidirectional long-short term memory (BiLSTM) network, which is designed to obtain more comprehensive and accurate predictions through bi-directional information capture, to identify which appliances are in operation. Finally, the prob-sparse attention mechanism is implemented to identify the multi-state processes of appliances by assigning different weights to the hidden layer units of the BiLSTM network. Throughout the network, convolution neural network (CNN) with different strides is introduced as down-sampling to ensure identical image sizes and concatenation layers are used for single-branch reconstruction of WD. In simulation studies, the model is trained and validated with low-frequency data in the UK-DALE and REDD datasets, and results show that compared with some state-of-the-art methods, the proposed method has a remarkable improvement in disaggregation performance.