Abstract: To solve the problem that the accuracy of railway passenger flow forecasting results was greatly affected by external objective environmental changes, and the accuracy of traditional passenger flow forecasting models was difficult to further improve effectively, this paper proposed an optimization method for railway passenger flow forecasting based on large language model. With the efficient text reading, content retrieval, and information integration capabilities of the DeepSeek-R1 model, the paper regularly screened key external events such as weather and special events that affect passenger flow, and added them as modeling key elements to the passenger flow forecasting model, effectively improved the accuracy of the model forecasting. Verified by actual railway passenger flow data from Beijing to Shanghai, the passenger flow forecasting model introduced with the large language model has reduced the average percentage absolute error by 5.5% compared to traditional time series forecasting models, and can effectively avoid forecasting anomalies caused by temporary external events. It has a good application effect in passenger flow forecasting work.