Abstract: Aiming at the problem of distributed photovoltaic electricity theft detection, a semi-supervised self-attention mechanism model based on multi-source data fusion(MDF-SSA) is constructed. Firstly, the PV cell simulation model is constructed, and multi-source features including PV power generation, light intensity, and temperature data are selected as inputs to minimize the impact of random fluctuations in PV power generation on electricity theft detection. Then the self-attention mechanism network and multilayer perceptron(MLP) are constructed to extract the features of multi-source data, and based on the idea of co-training, cross-co-training of the two networks is performed at the same time, the joint loss function is constructed and minimized, and the feature information in the labeled and unlabeled data are fully mined, and the detection accuracy is improved in the case of less-labeled data through semi-supervised learning. Finally, the optimal hyperparameters of the model are selected by the grid search method, and the PV electricity theft dataset is constructed using the real data from the Australian distribution network, and the model is validated by simulation experiments. The experimental results show that the proposed MDF-SSA model has better performance than other methods, with higher accuracy, recall rate, precision rate and F1-score and lower false positive rate.