Abstract: Non-intrusive load disaggregation (NILD) is able to decompose the aggregated energy into the equipment-level energy consumption, which is of great significance in the energy management, equipment fault detection and other fields. This paper proposes a NILD method based on deep learning for the low-frequency data. The method uses the multi-head probability sparse self-attention model in the natural language processing field to build a core disaggregation network. Taking the one-dimensional total power sequence as an input, it uses the convolution and pooling to extract the features. Combining with the positional encoding, it enhances the internal relationship between data in the sequence. By using the core disaggregation network the features are processed. By transposing the convolution and the full connection the features are mapped to generate the one-dimensional single electrical power. Finally, this paper uses the domestic appliance-level electricity (UK-Dale) to train and verify the model, and compares it with the existing three benchmark load disaggregation methods. The results show that the disaggregation performance of the proposed method has improved significantly.