Back Propagation neural network model (BP model) was applied to the Olympic air quality forecasting. By using MATLAB neural network toolbox, assembled air quality modeling forecasting system was combined with nearly real-time observations to fit BP neural network forecasting results. The measurements were made from 7 July to 26 August 2008 in the Peking University Health Science Center. Using these data, the performance of the BP neural network results is assessed. The results show that BP neural network can significantly improve the simulation results, with the average error rate decreasing by 34.7% and the correlation coefficient increasing by 39%. The advantage of BP neural network is more obvious when the original simulation results are poor. The sensitivity experiment of BP neural network sample size indicates that the number of samples is not the decisive factor in deciding fitting effect, and the stability of mapping relationship in samples plays a crucial role in raising the effect of prediction.