Hourly Prediction Model Of Ozone Concentrations Based On Ensemble Learning

The increasing of tropospheric ozone?O₃?concentrations has become one of the focused environmental problems all over the world.With the rapid development in an all-round manner,China also suffers from widespread O₃ pollution.Studies have shown,long-term exposures to high concentrations of ozone can cause serious damage to human health,animal and plant growth and even the ecological environment.However,the causes of O₃ pollution are complex and the governance is too difficult.In this situation,it has become more and more important to establish a timely,accurate prediction model of O₃ concentrations.Taking the Environmental Monitoring Station of Pudong New Area,Shanghai as an example,several relevant work has been done aiming at building an accurate and reliable hourly prediction model of O₃ concentrations in this paper.The main content includes the following:?1?This paper choosed the historical multi-factor variables related to O₃ from 2015to 2016 as research data.It is found that the accuracy of KNN algorithm can be improved greatly by giving the optimal weight to the neighboring samples.So the optimized KNN algorithm was used to fill the samples with low percentage of missing data in the original data,which guarantees the overall quality of the research data maximally.?2?By calculating the Pearson correlation coefficient between multiple factors and O₃ concentrations in the next 1²4 hours,the influence of multiple factors on O₃concentrations at different future time as well as their changing trends were explored.Moreover,the paper demonstrated the change rules of O₃ concentrations at three time scales of the month,week and hour,and explained the causes of those generally.?3?On basis of the original factors,prediction variables were extended by the periodic factors of O₃ concentrations and meteorological predictors.And the hourly prediction model was designed according to the characteristics of high autocorrelation between the concentrations of O₃ at adjacent time.With random forest,GBDT and XGBoost,the three ensemble learning methods,the continuous prediction of O₃concentrations in the next 1²4 hours was implemented respectively.The results show that the prediction effect of hourly prediction with the expanded data is obviously superior to the prediction effect by modeling directly with original factors data.What is more,during hourly prediction,the three ensemble learning algorithms all achieved high accuracy,and XGBoost has shown obvious advantages with R² above 0.88.?4?In order to improve the accuracy of O₃ concentrations prediction,this research attempts to use Stacking to combine random forest,GBDT and XGBoost,then apply the integrated model to hourly prediction.The results show that using Stacking for multi-model fusion can improve the accuracy of the hourly prediction model to a certain extent.The prediction result R² is above 0.89,which is 0.0085 higher than XGBoost by average.It has further improved the reliability of the prediction model.