A neural network model to predict spatiotemporal PM2.5 with FY-4A total precipitable water

Abstract

Heavy pollution weather has been proven by many to be very dangerous to human health and the normal functioning of society and the economy. PM2.5 is one of the causes of the formation of heavily polluted weather. At present, PM2.5 prediction results are only single station values or average values for small areas, but in our practical application large area gridded prediction values are the most valuable. So, we propose a FFT-ConvLSTM (Fast Fourier Transform ConvLSTM Network) model. The model first uses FFT to extract/analyze the common oscillation period of all input features and sets the time sliding window's size according to the period's length. Finally, not only high accuracy gridded prediction is achieved, but also long time prediction is realized by frame by frame output. In this study, air pollutants, meteorological elements and FY-4A water vapor data in the Beijing-Tianjin-Hebei region are used as input features of the model, among which FY-4A water vapor data solve the drawback that other monitoring devices cannot obtain high spatiotemporal resolution water vapor data. It is shown that the best common periods are 200 h, 133 h and 103 h for plains, mountains and plateaus in the Beijing-Tianjin-Hebei region, respectively. The RMSE of the FFT-ConvLSTM model is reduced by 5.4 μg/m3, 1.46 μg/m3 and 0.95 μg/m3 compared with the ConvLSTM model under the three different regions. The FFT-ConvLSTM model reduced the three-region mean RMSE by 3.69 μg/m3 and 5.2 μg/m3 compared to the MLR and CNN models, respectively.