PM2.5 is a major pollutant in the ambient air of open-pit mining, whose accurate prediction is significant for its removal and control design. In this study, a hybrid method was proposed for analyzing PM2.5 concentration. A field measurement of PM2.5 concentration was conducted at an operating open-pit mine in northern China. The data was cleaned to remove the outliers in the PM2.5 results. Gradient boosting machine (GBM) optimized by particle swarm optimization (PSO) was used for the regression and classification analysis. In terms of regression, different scenarios were designed to evaluate the effect of time interval for future predictions (5 min, 10 min, 20 min, 40 min, 1 h, and 2 h ahead) on the performance. The classification of PM2.5 concentration into ‘severe’ and ‘not severe’ was also analysed. The results are as follows: (i) A total of 37 data instances were detected to be outliers, and the regression performance was improved after data cleaning (from 0.902 to 0.937 on the training set and from 0.877 to 0.940 on the testing set). (ii) The percentage of training set percentage was determined to be 70%, and PSO performed well in optimizing the hyper-parameters of GBM, (iii) The regression was quite satisfactory with the correlation coefficient being larger than 0.9. The testing performance decreased with the increase in time interval. (iv) An average accuracy of 0.954 was achieved during the classification of PM2.5 concentration. The predicted PM2.5 concentration could work as a precursor for the heavy PM pollution around open-pit mining.