A deep learning method to repair atmospheric environmental quality data based on Gaussian diffusion

Abstract

Online monitoring data of atmospheric environmental quality often deviate or are missing, causing a great impact on regional atmospheric quality analysis. In this study, a deep learning method to repair atmospheric environmental quality data based on Gaussian diffusion and gate recurrent unit (GD-GRU) was developed to improve repair accuracy. A multi-source Gaussian diffusion model was developed to estimate PM2.5 based on the pollutant diffusion law and the data of 61 stations in Guilin. The root mean square error (RMSE) of the estimated and observed value was extracted as the error sequence. The error value was regarded as output of gate recurrent unit (GRU) with the inputs of weather and pollutant parameters. Missing data were calculated by Gaussian diffusion estimated value and the error predicted by GRU. The established GD-GRU model was applied to repair the long-sequence missing data. The analytical results indicated that the GD-GRU model had higher prediction accuracy of extreme values than Gaussian diffusion model and GRU model, because GD-GRU based on Gaussian diffusion can calculate the extreme value by simulating the diffusion and transmission mechanism. The established model predicted PM2.5 concentration in the next hours with an RMSE of 12.561, which was approximately 21.02% better, on average, than methods like autoregressive integrated moving average model (ARIMA), support vector regression (SVR), recurrent neural network (RNN), long short-term memory model (LSTM), and GRU. The established GD-GRU model demonstrated good performance on extreme values prediction and air quality data repair, thus providing a new method for air quality long-sequence missing data repair.