Aiming at the problems of poor prediction accuracy and overfitting of traditional time series prediction models in the inter-seasonal PM2.5 concentration prediction. In this paper, a parallel fusion model, the seasonal differential autoregressive sliding average model (SARIMA)-radial basis neural network (RBF) model, is proposed. Firstly, the seasonal factors of PM2.5 are extracted using the SARIMA model to analyze and predict the seasonal trend of PM2.5; then the effects of SO2, NO2 and CO on PM2.5 are extracted through the characteristic factor analysis, and the RBF neural network model, which has these three factors as inputs, is established to predict PM2.5; finally, the weighted fusion is used to achieve the accurate prediction of PM2.5 concentration by weighted fusion. Through the comparative test analysis of SARIMA model, RBF model and SARIMA-RBF parallel model, it is found that the prediction accuracy of SARIMA-RBF parallel model has been significantly improved, and the MAE of the parallel model reaches 3.23. This study proves that the SARIMA-RBF parallel model has a significant advantage in the prediction of PM2.5 concentration, which provides an important tool for the effective development of air pollution control strategies. This study demonstrates that the SARIMA-RBF concurrent model has significant advantages in the prediction of PM2.5 concentration, which provides important support for the effective formulation of air pollution control strategies.