Non-invasive Load Decomposition Model Based on Inception-SimAM-BiLSTM

Non-invasive load decomposition technology can obtain equipment-level power consumption information by disassembling household total load which can provide help for residents to improve the way of electricity consumption. To solve the problems of single feature extraction scale and low decomposition accuracy of current load decomposition models a sequence-to-sequence model based on Inception-SimAM (simple, parameter-free attention)-BiLSTM (bidirectional long short-term memory network) was proposed. This model includes an improved Inception module which incorporates Batch Normalization to avoid the problem of overfitting the model. At the same time, the SimAM mechanism is used to compute the three-dimensional attention weight of the characteristic sequence (considering the spatial and channel dimensions) to optimize the feature of the load sequence. Finally, the input sequence is decomposed by the BiLSTM. The results based on the REDD dataset indicate the proposed model extracts multi-scale features and feature concerns without SimAM through the improved Inception module, which significantly improves the decomposition accuracy.