Pollutant Forecasting Based On LSTM And WRF Model

Air pollution will cause harm to people’s body and psychology,and will also affect the ecological environment and economic development.Accurate pollutant forecasting is very important for air pollution prevention and control.At present,the WRF-Chem/WRF-CMAQ simulation system is commonly used to predict air quality,but its calculation is cumbersome,and the simulation accuracy greatly depends on the accuracy of the emission inventory of pollution sources and the reaction mechanism in many built-in modules,so the pollutant prediction effect is not ideal.For air quality monitoring stations in a fixed area,forecasting based on mathematical statistics or machine learning tends to be more effective.Meteorological conditions are a key factor influencing changes in pollutants,and it is not very accurate to use historical pollutants only to predict future pollutants.The numerical model WRF can simulate a variety of meteorological data and is more and more widely used in the meteorological field.This paper uses the meteorological data output by the WRF model and machine learning algorithms to predict six common pollutants PM_2.5,PM₁₀,SO₂,CO,NO₂ and O₃.The more accurate four-dimensional temporal air meteorological data obtained by the WRF model and the historical pollutant data were selected together with the characteristic variables.The characteristic variables screened were used as input parameters of the Machine learning models—Long Short Term Memory networks（LSTM）.Then six common pollutants were predicted respectively.The accuracy of the method for predicting multiple pollutants was studied by comparing with the measured data and the prediction results of the new generation atmospheric chemical model WRF-Chem.In this study,the numerical model WRF was used to simulate the weather process in Shaanxi in November and December 2021.The reanalysis data and Jinghe weather station data were used to compare the simulated high-altitude meteorological field and the hourly meteorological data of warm and humid winds on the ground.In order to screen out the characteristic variables related to pollutants as much as possible and improve the accuracy of forecasting,this paper proposes a method that combines Pearson correlation coefficient method with Lasso regression method to screen the characteristic variables.First set a lower Pearson coefficient threshold to screen out all possible related feature variables,and then filter out important feature quantities through Lasso regression to reduce redundancy.Combine the advantages of the two methods to select reliable feature variables as input parameters of the forecast model.The LSTM pollutant prediction model was established.The 24th,48th and 72nd hour single-step forecasting and seven-day multi-step forecasting for six pollutants were carried out.And the seven-day multi-step forecast had two methods:autoregressive prediction and single multi-step forecasting.In the single-step prediction of pollutants,the goodness of fit of the forecast results of the six pollutants in the 24th hour can reach more than 0.6.The effect of using the autoregressive prediction model for PM_2.5 and PM₁₀ in the next three days is better than that of the single-step forecasting model,while the single-step forecasting model has a better prediction effect on SO₂ in the next three days than the autoregressive forecasting model.The prediction effect of the two prediction models on the next three days of CO,NO₂ and O₃ pollutants is not much different.The single-time multi-step prediction model has stable forecasting effect and simple model,which can predict the change trend of pollutants.In addition,according to the cointegration theory,this paper constructs a multi-site joint forecasting model by taking advantage of the consistency of pollutant trends between several stations in Xi’an and its surrounding cities.Taking PM_2.5 and CO as examples,a joint forecast for the next 24 hours and seven consecutive days was carried out for a total of eight stations in Xi’an and several surrounding cities.The results show that the joint forecast model of PM_2.5 and CO has the best prediction effect on the site in Chang’an District.Combining with the forecasting of the two pollutants,the analysis of the joint forecast model can provide a forecast scheme for areas with consistent time series changes.