The Correction Of Atmospheric Visibility Prediction In Shanghai Based On LightGBM Framework

With the frequent occurrence of fog and haze weather in China in recent years,poor atmospheric visibility has occurred from time to time.The prediction of atmospheric visibility plays an important role in the prevention and control of urban air pollution and traffic safety.At present,the prediction accuracy of atmospheric visibility is not high.Therefore,it is of great social value to revise the prediction of atmospheric visibility in order to obtain better prediction accuracy.In this article,firstly,the characteristics of monthly,weekly and daily variations of three research areas in Shanghai were analyzed,and the linear correlation between atmospheric visibility and its lag terms were analyzed.Meanwhile,the monthly,weekly,daily and the lag terms features of atmospheric visibility were extracted as alternative features of the forecast correction models.Subsequently,forecast correction models of atmospheric visibility based on LightGBM framework were established in different research regions.Among them,the model building is divided into three steps.In the first step,according to the prediction accuracy difference of the original WRF mode,each research region was divided into three time periods to establish the model.In the second step,feature engineering was applied to the initial input features of the model,including feature synthesis and feature selection.The predicted values of horizontal and vertical components of wind speed in WRF mode at different altitudes were obtained by feature synthesis.By using variance selection method and feature contribution method based on GBDT algorithm,the optimal feature subset of the model was obtained as the final input features.The third step was to establish WRF mode atmospheric visibility prediction correction models based on LightGBM framework in different time periods.The optimal parameters of each model were obtained by grid search method.The stability of the models was verified by 10-fold cross validation method.Finally,the models were validated on each validation set.The results showed that the MAE between the predictions and the field data of the three research regions in Shanghai were all between 2097m and 331m,which increased by at least 2625m and47.2%compared with the original WRF mode.The RMSEs between the predictions and the field data were between 3164m and 4418m,which increased by at least 3239m and 45.1%compared with the original WRF mode,and the~2were between 0.76and 0.87,which increased by at least 0.56 compared with the original WRF mode.It can be concluded that the correction models of Shanghai atmospheric visibility prediction based on LightGBM framework had good prediction accuracy and can significantly improve the prediction accuracy of the original WRF mode.