Non-intrusive residential load disaggregation is of great significance for mining the energy demand on the user side. However

the current power decomposition model does not converge well and the inference cycle is computationally intensive. A non-intrusive load decomposition model based on sequence-to-subsequence and subtask gated networks(SGN )is investigated in the paper. Firstly

a sequence-to-subsequence approach is used to construct a sub-task gating network

combining a power decomposition regressiontaskwithanappliancestateclassificationtaskto achieve a mapping from the main power sequence to the target appliance sub- series. Then a channel attention module and a spatial attention module are added to improve the feature extraction capability of the model. Experimental results based on the UK-dale da taset show that the method not only reduces the difficulty of convergence of the model and the computational effort of the inference cycle

but also significantly improves the accuracy of the decomposition.