PM2.5 is a type of air pollutant that can cause respiratory and cardiovascular problems. Precise PM2.5 (µg/m 3 ) concentration prediction may help reduce health concerns and provide early warnings. To better understand the air pollution, a number of approaches have been presented for predicting PM2.5 concentrations. Previous research used deep learning models for hourly predictions of air pollutant due to their success in pattern recognition, however, these models were unsuitable for multi-site, long-term predictions, particularly in regards to the correlation between pollutants and meteorological data. This article proposes the combine deep network (CombineDeepNet), which combines multiple deep network, including a bidirectional long short-term memory (BiLSTM), bidirectional gated recurrent unit (BiGRU), and a shallow model represented by fully connected layers (FC), to create a hybrid forecasting system. It can effectively capture the complex relationships between air pollutants and various influencing factors to predict hourly PM2.5 concentrations in multiple monitoring sites based in China. The best root mean square error (RMSE) achieved was 22.0 µg/m 3 (long-term) and 6.2 µg/m 3 (short-term), with mean absolute error (MAE) values of 3.4 µg/m 3 (long-term) and 2.2 µg/m 3 (short-term). Additionally, the correlation coefficient (R 2 ) reached 0.96 (longterm) and 0.83 (short-term) across six monitoring sites. These results demonstrate that CombineDeepNet enhances prediction accuracy compared to popular deep learning methods. Therefore, CombineDeepNet proves to be a important framework for predicting PM2.5 concentration.