Abstract: In this study, a multi-perturbation air quality ensemble prediction system is developed, based on the Nested Air Quality Prediction Modeling System (NAQPMS), by the Monte Carlo simulation method. The system is used to predict the PM2.5 concentration of the "2+26" cities in Beijing–Tianjin–Hebei. Further, the period of the experiment is from September to December 2017, with a horizontal resolution of 15 km. This study found that the method of ensemble mean after selecting ensemble samples can significantly improve the accuracy of PM2.5 forecasting and greatly reduce the forecasting bias in the system, based on the Monte Carlo ensemble forecasting system. Compared with the ensemble mean method of all the samples, this method reduces the RMSE of PM2.5 forecast from 58.0 μg m⁻³ to 34.7 μg m⁻³, increasing the fraction of prediction within a factor of two of observations from 67% to 87%. Furthermore, the ensemble mean method after selecting ensemble samples is better than the ensemble mean method for the overall forecast of each pollution level. The results of this study can provide a reference for improving the impact of the urban PM2.5 forecast and reducing the bias of the PM2.5 forecast.