Abstract: Aimed at the problem of single feature extraction method in deep learning in non-intrusive load decomposition, a method of spatiotemporal feature fusion attention is proposed. On the one hand, five parallel convolutions were used to learn multi-scale spatial features from the input sequence. On the other hand, BiLSTM was used to learn temporal features from the input sequence. The learned temporal and spatial features were cascaded and fed into a convolution attention network module for weighted fusion learning of spatiotemporal features, thereby improving the accuracy of non-intrusive load decomposition. The simulation results show that the proposed method outperforms the existing deep learning methods in both seen and unseen scenes, which is of great significance to improve the accuracy of load decomposition.