Prediction of PM2.5 in Hefei Based on a Hybrid CNN-GRU Model

Abstract

The world has been affected by air quality problems, especially PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> pollution for a long time. The existing prediction methods based on deep learning often pay attention to the influence of time series information on the prediction of PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> concentration, but ignore the importance of spatial information, the prediction effect is unstable and the prediction accuracy of PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> peak value is not high. In this paper, a method of combining Convolution Neural Network with Gated Recurrent Unit (CNN-GRU) is proposed to predict PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> concentration by using current and past PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> concentration data, meteorological data and time stamp data, considering the comprehensive influence of time series information and spatial information. The model consists of three parts: CNN structure, GRU network and fully connected network. Among them, the CNN structure can automatically extract and integrate the local trends and spatial correlation features of multi-site multimodal air quality data, and the GRU network can capture the long-term dependent features of time series. The combination of the first two parts can automatically extract the spatial and temporal features of air quality data and improve the accuracy of PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> prediction. Finally, the fully connected network outputs the final prediction result of PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> . This study selects three monitoring stations from Hefei area as the research object. The prediction effect of hybrid CNN-GRU model is compared with that of traditional RNN, LSTM and GRU models, and the prediction performance of different models under different prediction steps is evaluated. The results show that the hybrid CNN-GRU model proposed in this paper has better prediction performance and stability than other models due to the introduction of spatial features.