Research of PM2.5 Prediction System Based on CNNs-GRU in Wuxi Urban Area

Abstract

Aiming at the serious pollution situation and lack of effective prediction methods in Wuxi urban area, based on convolutional neural network (CNN) and gated recurrent unit (GRU), this paper proposes a PM2.5 prediction model that can automatically extract spatiotemporal features of multi-station and multimodal air quality data, and build a PM2.5 prediction system based on this model as well. The system model firstly takes multiple two-dimensional (2D) matrices constructed with time series of the air quality factors and weather factors from different monitoring stations in Wuxi urban area as input, automatically extracts and fuses the local variation trends and spatial correlation features of multi-station and multimodal data with CNNs structure. The results from the CNNs are input to the GRU network to further capture the long-term dependence feature of air quality data. Then, a fully connected network taking the spatiotemporal features as input is used to predict the PM2.5 concentration for the next 6 hours in Wuxi urban area. The PM2.5 prediction system based on CNNs-GRU model is tested on the real data set provided by Wuxi Environmental Protection Bureau. On the two test sets in January and June, the prediction accuracy of the PM2.5 prediction system reached 76.902% and 70.053% respectively, which is better than the comparative models. Finally, the prediction system based on the optimal CNNs-GRU model and real-time data obtained by crawlers, predicts the real-time PM2.5 concentration for the next 6 hours, and visualizes the prediction results on the Web through Echarts technology. It can provide valuable reference for citizens' travel, prevention and control of air pollution in Wuxi urban area.