| At present,the ash pollution with PM2.5 as the main pollutant has become a serious environmental problem in China.Due to the uneven distribution of monitoring stations and the lack of historical data,the available PM2.5 concentration data obtained by stations was discontinuous in time and space.This problem has led to the lack of reliable data for the study of PM2.5 concentration distribution,change,prediction and so on.Therefore,in order to obtain high-quality data of spatiotemporal continuity,many studies have estimated PM2.5 concentration by remote sensing,statistics,machine learning and other methods.There are still some problems in current research.For example,the accuracy of the estimation results needs to be improved.The estimation method of remote sensing inversion relies on the quality of remote sensing images.The time and space accuracy of the estimation is low,and it is difficult to retrieve the historical data of PM2.5 concentration.And the model is applicable only to specific areas or time frame.The machine learning model was only applied to small-area regions in the previous studies.And the studies often neglected the temporal and spatial continuity of PM2.5 concentration distribution,the periodic characteristics of PM2.5 concentration variation on different time scales,and the impact of landscape pattern on PM2.5concentration distribution.Ensemble learning is a combinatorial optimization algorithm based on traditional learning machine.It has achieved better performance than single learning machines in many fields including speech recognition and image recognition.In order to improve the accuracy and universality of the estimation model and solve the problems existing in the study of PM2.5 concentration estimation,this paper used ensemble learning method to estimate the concentration of PM2.5.Five traditional machine learning algorithms were selected based on the accuracy of a single model.The diversity and accuracy of individual machine combinations were evaluated by deviation-variance-covariance decomposition method.On this basis,8 kinds of ensemble model were established by Stacking ensemble method and using linear regression as meta-regressor.In the model training stage,the traditional machine learning feature library of PM2.5 has been optimized.We added the temporal and spatial characteristics of PM2.5 concentration to simulate the distribution of PM2.5 concentration around the specified location.The periodicity of the PM2.5 concentration has been considered in different time scales.At the same time,the relationship between landscape pattern and PM2.5 concentration was concerned,and the impact was characterized by landscape indices such as Patch Density.In this paper,model training and accuracy evaluation were carried out based on the optimized feature library.Finally,we used the Stacking ensemble model to estimate the PM2.5 concentration per hour in Beijing on March 23,2017.The experimental results showed that the Stacking ensemble model with K-nearest neighbor,BP neural network,extreme tree and XGBoost as level-1 regressor and linear regression as meta-regressor?KNN-BP-ET-XGB—LR?had the best effect.The2 for training set and test set was 0.9890 and 0.8812 respectively,which were higher than traditional machine learning methods and statistical regression model.The model's effect and applicable space-time range are also superior to other studies.The effect of the ensemble model was not directly related to the number of individual machine.The ensemble model combined by different kinds of individual models was better than those of the same kinds of individual models.The Bias-Variance-Covariance Decomposition showed the biggest difference between KNN model and other models.It is suitable for coupling with other models.This index could reflect the effect of the final ensemble model to some extent,but it could not be used as an only index to evaluate model's performance.In this paper,the new space-time,meteorological and periodic characteristics make the model more accurate to obtain the factors affecting the change of PM2.5.When these three characteristics are deleted,the2 of the integrated model is reduced from 0.8812 to 0.7344. |
PDF Full Text Request