基于深度学习的大气细颗粒物污染时空预报

Spatio-temporal Forecast of Ambient Fine Particulate Matter Pollution Based on Deep Learning

为尽可能减少PM2.5重污染带来的环境和健康危害,对其进行及时准确的预报预警显得十分关键。文章使用WRF气象模型和CMAQ空气质量模型输出的2015年1-12月逐小时的中国地区27 km×27 km网格化气象场、污染排放和细颗粒物模拟浓度数据,训练深度学习预报模型。研究结果显示,深度学习预报结果与CMAQ模拟结果在测试集上的均方误差为10.52μg/m3,优于已有的大部分其他研究,深度学习在重点城市的PM2.5预测浓度趋势与CMAQ基本一致,其RMSE为28.82μg/m3,整体空间分布也具有较好的一致性,可以准确重现CMAQ模拟结果。该研究进一步使用国控点实际观测数据对训练完成的深度学习预报模型进行优化,以减少CMAQ理论模型所带来的固有误差。优化后的深度学习模型预报结果在总体上更接近实际观测的真实结果。此外,深度学习模型的预报速度在相同CPU环境下比CMAQ模型提升93倍,在GPU加速环境下提升高达465倍,可以做到快速实时的浓度预报响应。该研究构建的耦合数值模型和实际观测数据的深度学习预报技术,能够快速准确地给出大气细颗粒物浓度的未来时空分布预报结果,对深度学习技术在实际预报业务中的应用具有重要的借鉴价值。

It is essential to make timely and accurate forecasting as well as early warning of PM2.5 to minimize the environmental and health damage caused by pollution of PM2.5.In the paper,the WRF and CMAQ models were used to obtain27 km×27 km hourly grid data in terms of meteorological field,pollutant emissions and PM2.5 concentration values in China from Jan.to Dec.of 2015 for training the deep learning the forecasting model,i.e.,Conv-LSTM(convolutional long-short-term memory).The result showed that RMSE between deep learning and CMAQ on test dataset was 10.52μg/m3,being better than the results of most existing studies.The trend of PM2.5 concentration in main cities of China predicted by deep learning model was nearly similar to the results of CMAQ and the RMSE was 28.82μg/m3;the overall spatial distribution also had a good consistency;and CMAQ simulation results could be reproduced as well.Furthermore,the observed datasets from the National Stations were used to optimize the deep learning model and reduce the errors from CMAQ.The values of PM2.5 concentration predicted by optimized deep learning model were generally closer to real observed results.Compared with CMAQ,the speed of deep learning forecasting model increased by 93 times faster than that using CPU and 465 times faster than that using GPU acceleration.In conclusion,the deep learning forecasting model coupled with numerical model and observed data developed in the study can quickly and accurately predict the future spatial-temporal distribution of fine particulate matter concentration,and hopefully,it would be useful in the practical PM2.5 forecasting by use of deep learning technology.