Multi-hour and multi-site air quality index forecasting in Beijing using CNN, LSTM, CNN-LSTM, and spatiotemporal clustering

Effective air quality forecasting models are helpful for timely prevention and control of air pollution. However, the spatiotemporal distribution characteristics of air quality have not been fully considered in previous model development. This study attempts to establish a multi-time, multi-site forecasting model of Beijing's air quality by using deep learning network models based on spatiotemporal clustering and to compare them with a backpropagation neural network (BPNN). For the overall forecasting, the performances in next-hour forecasting were ranked in ascending order of the BPNN, the convolutional neural network (CNN), the long short-term memory (LSTM) model, and the CNN-LSTM, with the LSTM as the optimal model in the multiple-hour forecasting. The performance of the seasonal forecasting was not significantly improved compared to the overall forecasting. For the spatial clustering-based forecasting, cluster 2 forecasting generally outperforms cluster 1 and the overall forecasting. Overall, either the seasonal or the spatial clustering-based forecasting is more suitable for the improvement of the forecasting in a certain season or cluster. In terms of model type, both the CNN-LSTM and the LSTM generally have better performance than the CNN and the BPNN.

Automated summary

This study established a multi-time, multi-site forecasting model of Beijing's air quality using deep learning network models based on spatiotemporal clustering. The LSTM model was found to be the optimal model for multiple-hour forecasting. Seasonal forecasting did not show significant improvement.