Prediction of PM2.5 Hour Concentration Based on U-net Neural Network

Most of the previous PM2.5 prediction models present unsatisfactory performance in several aspects, including predicting accuracy and generalization ability, especially in case of the sudden change in the value of PM2.5 situation. Therefore, we propose a method based on the U-net neural network to predict the hourly PM2.5 concentration value on the research area, attempting to improve the prediction performance. The proposed model includes two major steps. First, based on the inverse distance interpolation of historical wind field data, discrete station PM2.5 values are interpolated into a PM2.5 grid map; second, the U-net neural network is applied to train the prepared spatiotemporal grid data and make predictions. The model can use the PM2.5 concentration values of the grid map extracted at different time stamps for the PM2.5 prediction. The PM2.5 concentration values at all locations in the research region can be achieved. Specifically, the prediction accuracy and the generalization ability of the model in case of sudden changes are revealed. Experimental results indicate that the proposed method has a 10% improvement in the prediction accuracy of PM2.5 concentration values in the case of sudden change. © 2020 Peking University.