Due to the diverse grid-connection methods and flexible operation modes of distributed photovoltaic system, electricity theft behaviors are highly covert, making it urgent to develop efficient detection methods. Aiming at the problems of weak generalization ability and low detection accuracy of existing supervised classification and unsupervised regression electricity theft detection models, a hybrid semi-supervised model combining convolutional neural network and Transformer is proposed for electricity theft detection. Specifically, the convolutional neural network (CNN) model integrates the bidirectional long short-term memory network and the attention mechanism, and employs an improved whale optimization algorithm (incorporating optimization strategies such as Tent chaotic mapping and nonlinear convergence factors) to search for optimal hyperparameters. The Transformer model is optimized through a multi-scale perception layer and a linear mapping layer. Experimental results demonstrate that, in scenarios with a small amount of labeled data and a large amount of unlabeled data. Compared with conventional methods, the accuracy of the proposed method is improved by 4.98% and 0.44%, respectively, and the area under the curve is increased by 3.12% and 0.74%, respectively. This research provides new ideas for anti-electricity theft technologies in power system and facilitates the sustainable development of clean energy.