Gated Recurrent Unit Coupled with Projection to Model Plane Imputation for the PM2.5 Prediction for Guangzhou City, China

Air pollution is generating serious health issues as well as threats to our natural ecosystem. Accurate prediction of PM2.5 can help taking preventive measures for reducing air pollution. The periodic pattern of PM2.5 can be modeled with recurrent neural networks to predict air quality. To the best of the author’s knowledge, very limited work has been conducted on the coupling of missing value imputation methods with gated recurrent unit (GRU) for the prediction of PM2.5 concentration of Guangzhou City, China. This paper proposes the combination of project to model plane (PMP) with GRU for the superior prediction performance of PM2.5 concentration of Guangzhou City, China. Initially, outperforming the missing value imputation method PMP is proposed for air quality data under consideration by making a comparison study on various methods such as KDR, TSR, IA, NIPALS, DA, and PMP. Secondly, it presents GRU in combination with PMP to show its superiority on other machine learning techniques such as LSSVM and two other RNN variants, LSTM and Bi-LSTM. For this study, data for Guangzhou City were collected from China’s governmental air quality website. Data contained daily values of PM2.5, PM10, O3, SOx, NOx, and CO. This study has employed RMSE, MAPE, and MEDAE as model prediction performance criteria. Comparison of prediction performance criteria on the test data showed GRU in combination with PMP has outperformed the LSSVM and other RNN variants LSTM and Bi-LSTM for Guangzhou City, China. In comparison with prediction performance of LSSVM, GRU improved the prediction performance on test data by 40.9% RMSE, 48.5% MAPE, and 50.4% MEDAE.