Abstract: To improve the forecast accuracy of PM_2.5 concentration in heavily polluted regions of China, ensemble forecasts were built by mean ensemble, weighted ensemble, multiple linear regression ensemble and back propagation artificial neural networks ensemble, respectively, based on four numerical air quality models developed and maintained by national or regional environmental metrological service centers. A best multi-model ensemble forecast was established based on real-time evaluations of performances of single numerical models and ensemble methods. Through evaluation of the forecast results during 2015-2016, compared with single numerical air quality forecast models, improvements on forecast biases due to mean and weighted ensembles were limited, but multiple linear regression, back propagation artificial neural networks and best ensembles could largely reduce the forecast biases. The NMB and RMSE values between best ensemble forecast and observation were from -10% to 10% and from 10 to 70 μg/m ³, respectively. Best ensemble showed strong correlation with observations at more sites compared with other ensemble methods, but also underestimated PM_2.5 concentrations in high pollution level. During the pollution process occurred in Jing-Jin-Ji region from February 25 to March 4, 2018, best ensemble had the ability to forecast the trend and magnitude of PM_2.5 concentrations. In three representative cities of Beijing, Shijiazhuang and Zhengzhou, the NMB and R values between best ensemble and observations varied from 26% to 4% and from 0.49 to 0.77, respectively. The TS scores of best ensemble for mild and moderate pollution ranged from 0.39 to 0.73, and that of severe and above pollution ranged from 0.13 to 0.30. These indicate that best ensemble can provide a strong objective reference to forecaster, but its forecast ability of peak values needs to be further improved.