Study Of Haze Feature Extraction And Pollution Grade Recognition And Early-Warning Technologies

In recent years,the study of problems causing haze becomes popular.The Beijing-Tianjin-Hebei region of China has been seriously polluted by PM2.5.In order to improve the detection accuracy of haze early-warning and reduce the harm to human production and life,this paper aims to study the methods of haze feature extraction and pollution grade prediction at Beijing-Tianjin-Hebei area.However,the main issue of current haze feature extraction method is only based on the correlation between features,which may result in large prediction error.To solve it,this paper proposes a fuse feature extraction method,such as a high-level feature extracted by Deep Belief Networks(DBN),and a low-level feature extracted by traditional method.The factors causing haze are analyzed deeply.The characteristic factors which lead to haze are synthetically extracted from both meteorological and non-meteorological factors.The low-level features are extracted by tree-based low-level feature extraction method,and the high-level features are extracted by DBN.Then the high-level features and low-level features are fused as input features for training the prediction model.Compared with the classical principal component analysis method and the feature extraction method based on tree model,the feature extracted by the fuse feature extraction method presented in this paper is more effective.Currently,there are mainly two kinds of haze prediction methods:traditional numerical and statistical prediction and artificial neural network prediction.The former has poor haze prediction results based on a set of non-linear data,which makes it hard to obtain satisfactory results.The latter is prone to complicated parameter adjustment,slow convergence speed and easy to fall into local minimization.To solve these mentioned issues,this paper proposes a new haze prediction method which combines deep confidence network with XGBoost algorithm.This method connects the last layer of deep confidence network to the XGBoost model together with the lower feature layer to train a model with a training dataset.This paper has proposed a PM2.5 pollution classification index,classified the predicted results,and optimized the early-warning results through expert experience knowledge.This method performs better than the support vector regression method,which performs the best in the literature and the widely used BP neural network method.