Research On Shanghai Visibility Prediction Model Based On Multi-source Data And XGBoost Algorithm

The factors affecting visibility are numerous and the mechanism is complex,which poses a great challenge to the visibility forecast.It is very meaningful to accurately forecast visibility for ensuring traffic safety and improving trip quality.Shanghai,as one of the largest cities in China,is one of the most important transportation hubs in China.It has important practical guiding significance for public safety management to study on an accurately forecast visibility system for this megacity.In the aspect of visibility forecast research,the numerical prediction models that are widely used ignore the non-linear relationship between forecast factors and visibility,while the research on visibility prediction using machine learning algorithm mostly relies on the data from observation station and lacks the mining of products generated from numerical prediction model.In addition,the previous researches pay more attention to the accuracy of prediction,but lack of exploration and explanation of the model's internal mechanism.To solve these problems,this study presents a visibility prediction method based on multi-source data and machine learning algorithm.The main contents and conclusion are as follows:1)Establishing the prediction model based on multi-source data and machine learning algorithm.In this study,combined three kind of climatic data from observation,EC-thin data and WRF?Weather Research Forecasting?model,we establish the prediction model of visibility based on XGBoost?eXtreme Gradient Boosting?algorithm,then using this model,we forecasted the visibility in future 24 hours for 11stations of Shanghai.Results showed that this model has a high accuracy and was more accurate than the WRF model which is a kind of the numerical prediction model.In detail,this model gets² at 60.2%and accuracy rate of classification at 81%.2)The importance analysis of features which affects visibility forecast.In this study,we have obtained the features that have significant influence on the improvement of model accuracy through calculated the importance of each prediction feature.This study found that pollutants concentration,wind speed,wind direction and relative humidity have great influence on the accuracy of model.In addition,EC-thin forecast data has a greater impact compare with WRF forecast data and observation data.3)The relationship between feature contributions and model.Calculating feature contributions which influenced the value of forecasting based on decision routine estimation.Results showed that visibility at forecast time,visibility and concentrations of PM_2.5 forecasted by WRF have significance contribution to visibility.With the increase of forecast visibility,the contribution of observation data,WRF forecast data and EC-thin forecast data generally change from small to large,from negative to positive.In addition,forecast data?WRF forecast data and EC-thin forecast data?have important contribution to the summation of feature contributions.