Timely and accurate air quality forecasting is of great significance for prevention and mitigation of air pollution. However, most of the previous forecasting models only considered either temporal or spatial prediction. The effects of different inputs, model structures and leading hours on the model performance were still less investigated. In this study, three kinds of deep learning neural networks (the long short-term memory (LSTM), the convolutional neural networks (CNN), and CNN-LSTM) were developed for 1- to 24-h forecasting of spatiotemporal PM_2.5 concentrations. The back propagation neural network (BPNN) was employed as the control model. Three types of input data (previous PM_2.5 concentrations plus meteorological data, air quality data and the six most correlated variables, respectively) at multiple stations in Beijing from 2015 to 2016 were employed for model development. The results showed that PM_2.5 concentrations in the past hours, hybrid air quality indexes or other pollutant concentrations, and wind speed can be used as good candidate predictors. For the 1-h forecasting, the LSTM and the CNN-LSTM with the most correlated variables were the best models with top 2 performance. For the 2- to 24-h forecasting, the LSTM was the optimal model for forecasting over 12 h while the CNN-LSTM was generally better for forecasting within 12 h. This study demonstrated that deep learning outperformed shallow learning models in hourly PM_2.5 forecasting and suggested that optimizing the input and structure of the model helped to improve the performance of multiple-hour PM_2.5 forecasting.