To achieve accurate load awareness and efficient device control, this study proposes a novel non-intrusive load decomposition （NILD） method by combining multi-modal feature selection techniques with an improved time convolutional network to enhance identification accuracy and efficiency. In this method, the load multi-scale maximum overlapping discrete wavelet transform results are introduced as input features, and the feature is combined with the mRMR algorithm to ensure the effective integration of multi-modal features. At the same time, the adaptive receptive field attention mechanism （RFAM） is added to the TCN method,improving the ability to extract features at different time scales. Finally, the algorithm is validated on datasets such as PLAID. The results show that the proposed method can achieve more accurate and efficient load decomposition with fewer features,which is higher than the regression accuracy of the traditional BiLSTM model of 5.31%,and the detection rate of new energy household equipment such as photovoltaic power has also reached over 96.6%.

electric load management  /  wavelet transforms  /  feature selection  /  non-intrusive load decomposition  /  receptive field attention mechanism  /  temporal convolutional network