Nowadays，non-intrusive load disaggregation techniques face two major challenges.First，it is difficult to effectively extract the power characteristics of multi-state appliances in low-power states.Second，the generalization capability of disaggregation models is insufficient.To address these two challenges，an improved Unet++-based approach for multi-state appliance load disaggregation is proposed.First，within the encoder-decoder framework，a parallel-structured encoder is adopted to enhance the parsing capability of complex power signals，while skip connections ensure that the decoder can accurately reconstruct the original signal，thus improving the refinement of the disaggregation.Second，a bidirectional long short-term memory （BiLSTM） module is introduced to capture long-term dependencies in time series，enhancing the learning and prediction capability of the model.Experimental results show that the proposed model accurately identifies and disaggregates multi-state appliances on both the UK domestic appliance-level electricity dataset （UK-DALE ） and the reference energy disaggregation dataset （REDD）.In terms of mean absolute error，the proposed model demonstrates superior performance，and results obtained from tests on publicly available datasets indicate that its performance is better than that of existing methods.