Convolution Recurrent Neural Networks for Short-Term Prediction of Atmospheric Sensing Data

A deep neural network based prediction model proved to be more efficient in a short-term PM 2.5 pollution prediction system than the mathematical statistics models that were applied. In this paper, we use the spatial-temporal characteristics of atmospheric sensing data with a deep neural network prediction model for short-term PM 2.5 pollution prediction in Fukuoka, Japan. To facilitate this, we introduce a convolution recurrent neural network (CRNN) based on a convolution neural network (CNN) and a long short-term memory (LSTM) that is suitable for short-term prediction with complex space-time series. We use the atmospheric sensing data of the transboundary air pollution data from 33 coastal cities in China and Fukuokas environmental monitoring data from 2015 to 2017 to train and test the CRNN. The results show that CRNN using the spacetime series is superior to a traditional neural network that only used time series inputs and the regression statistics prediction model used in Fukuoka. This study provides a new solution for predicting PM 2.5 pollution with atmospheric sensing data.