Abstract: Based on the epidemic data of novel coronavirus pneumonia (COVID-19) in Tokyo, Japan, in 2022, this study proposes a hybrid forecasting framework that integrates the dynamically modified extended SEIR (susceptible-exposed-infectious-recovered) model and the LSTM (long-term and short-term memory network) deep learning network to improve the prediction accuracy of epidemic trend. Based on the Tokyo epidemic data from September 1 to October 12, 2022, this study introduces an LSTM network to dynamically correct the residuals between the extended SEIR prediction values and the real data. Through the long short-term memory mechanism, nonlinear patterns in historical data are learned to achieve real-time compensation for prediction errors. The research results indicate that the optimized hybrid model is highly consistent with actual data in predicting the evolution of the Tokyo epidemic, and its peak prediction ability is significantly better than the single extended SEIR model.