基于深度学习方法对兰州市ρ(PM_(2.5))的模拟

A simulation study ofρ(PM_(2.5))in Lanzhou City proper based on deep learning methods

为准确、快速地模拟ρ(PM_(2.5)),构建深度学习模型:深度神经网络(DNN)、长短期记忆递归神经网络(LSTM)、卷积神经网络,用兰州市气象站监测数据、大气污染物排放清单以及环境空气质量监测站点的常规污染物监测数据,对兰州市逐小时ρ(PM_(2.5))进行模拟.结果表明,年际尺度上,3种模型中DNN的效果最好,LSTM对实测值较大的数据模拟效果比其他模型更好.季节尺度上,划分不同季节进行模拟的效果优于使用全年数据的模拟效果.3种模型中表现最好的是LSTM模型,整体表现为春、夏、秋季的模拟效果较好,冬季模拟效果较差.

In order to simulate theρ(PM_(2.5)),an important pollutant,more accurately and quickly,three deep learning models were constructed,i.e.deep neural network(DNN),long and short-term memory recurrent neural network(LSTM),and convolutional neural network;at the same time,the monitoring da-ta of Lanzhou City meteorological station,inventory of air pollutant emissions and the conventional state-controlled stations of ambient air quality monitoring were used.The pollutant monitoring data,the hour-by-hourρ(PM_(2.5))in Lanzhou City proper,were simulated.The results showed that the DNN model was the most effective of the three models at the yearly scale,and the LSTM model performed better than the other two in simulating data,with larger measured values.On the seasonal scale,the simulation of differ-ent seasons was better than simulation using year-round data,and the LSTM model performed better than the other three models.On the whole,the simulation results were better in spring,summer and fall,but worse in winter.