基于深度学习模型的广州市大气PM_(2.5)和PM_(10)浓度预测

PREDICTION OF PM_(2.5) AND PM_(10) CONCENTRATION IN GUANGZHOU BASED ON DEEP LEARNING MODEL

精准预测大气污染颗粒物PM_(2.5)、PM_(10)浓度能为大气污染防治提供科学依据,但目前较多PM;和PM;浓度预测在缺少污染源排放清单和能见度数据时,预测精度不高。而目前深度学习模型应用于PM;和PM;浓度预测的研究还鲜见报道。基于广州市2015年6月1日—2018年1月10日的空气质量和气象监测历史数据,分别构建了随机森林模型(RF)、XGBoost模型2种传统的机器学习模型和长短时记忆网络(LSTM)、门控循环单元网络(GRU)2种深度学习模型,并对广州市的PM_(2.5)、PM_(10)日均浓度值进行预测。结果表明:在缺少污染源排放清单和能见度数据时,4种模型也能较好地预测PM_(2.5)、PM_(10)日均浓度。根据MSE、RMSE、MAPE、MAE和R;等评价指标,对4个模型的PM_(2.5)、PM_(10)预测效果进行测评,得出深度学习GRU模型预测效果均为最佳,RF模型的预测结果均为最差。相比目前研究及应用较多的RF模型、XGBoost模型、LSTM模型,基于深度学习的GRU模型能更好地预测PM_(2.5)、PM_(10)浓度。

Precisely predicting the concentration of PM_(2.5) and PM_(10) in air pollution can provide a scientific basis for the prevention and control of air pollution. However, in the absence of pollution source emission inventory and visibility data, the prediction accuracy of the existing PM_(2.5) and PM_(10) concentration prediction methods are not high. In addition, it is rarely reported that the current deep learning models have been applied successively to PM;and PM;concentration prediction research. Based on the historical air quality monitoring data and weather monitoring historical data in Guangzhou from June 1, 2015 to January 10, 2018, two traditional machine learning models(random forest model(RF) and XGBoost model) and two deep learning models(short-long-term memory network(LSTM) and gated recurrent unit network(GRU) were constructed respectively, to predict the daily average concentration of PM_(2.5) and PM_(10) in Guangzhou. The results showed that the four models could also well predict the daily average concentration of PM;and PM;in the absence of pollution source emission inventory and visibility data. According to the evaluation metrics, i.e., MSE, RMSE, MAPE, MAE, and R^(2), the PM_(2.5) and PM_(10) prediction effects of the four models were evaluated. The results indicated that the prediction effect of the deep GRU model was the best and the prediction results of the RF model were the worst. Compared with the commonly used RF model, XGBoost model, and LSTM model, the GRU model based on deep learning could better predict PM_(2.5) and PM_(10) concentration.