Abstract:To decrease the forecast uncertainties of single models and improve the refinement of multi-model ensemble air quality forecast system, the gridding observed pollutant concentration with resolution of 0.25°×0.25° was firstly established by using Cressman interpolation method. Then, combined with four numerical air quality forecast models, the mean, weighted and multiple linear regression ensembles were established in each grid, respectively. Finally, based on the evaluation results of single models and ensemble methods in previous 50 days, an optimal ensemble was established. The evaluation results of PM2.5 concentrations during a heavy pollution process in 19-22 December 2018 showed that in the case of heavy pollution, the NMB values between the optimal ensemble forecast and observations could also be maintained between -20% and 40%. And the forecast coverage area with good and above pollution by the optimal ensemble was closer to observation than those of single models. During the whole process, the NMB, root mean squared error (RMSE) and R values between forecasted PM2.5 concentrations by the optimal ensemble and observation were from -20% to 20%, from 35 to 75 μg·m-3 and higher than 0.4, respectively, in most polluted areas. Among all single models and ensemble methods, number of girds over China with high total scores was the largest in optimal ensemble. In the eight cities located in the most polluted region, the average onset and end times of the pollution process by optimal ensemble forecast was 1.8 and 6.9 h earlier than observation, respectively. Therefore, we propose that pollutant concentrations retrieval by satellite and surface observation should be fused to improve the refinement of gridding observed pollutant concentrations. And the methods of scale reduction, subjective and objective fusion and rolling correction should be used to further improve the forecast accuracy of gridding multi-model ensemble air quality forecast.