The accuracy of anomaly detection in current distribution network load data anomaly detection is low, a load anomaly detection method for distribution network is proposed, which combines the improved long short-term memory network with variational auto-encoder. Long short-term memory network is optimized through residual structure to improve feature learning ability, and the optimized long short-term memory network replaces the BP neural network layer (encoding and decoding) of the variational auto-encoder, which can better obtain the time correlation of load data. By comparing with conventional testing methods, the superiority of the proposed detection method has been verified. The results indicate that, compared to conventional load data anomaly detection methods, the proposed method has better detection accuracy, the accuracy rate of anomaly detection reaches 97.30%, compared to not introducing residual structure, it has increased by 1.70%, improved by 7.00% compared to the LSTM model，improved by 4.80% compared to the PSO-PFCM model, which can provide a certain reference for the development of distribution network automation.