Research On PM_2.5 Concentration Prediction Based On Bidirectional Recurrent Neural Network

In recent years,the rapid development of urbanization and industrialization in China has brought tremendous damage to the environment,and the ensuing air pollution has seriously affected people’s daily life and health.Therefore,it is important to take reasonable measures to prevent and control air pollution.Accurate prediction of fine particulate matter(PM_2.5)change law can provide theoretical support for the prevention and control of air pollution,so as to realize effective prevention and control of air pollution.However,with the rapid development of Internet of Things technology,the traditional PM_2.5 prediction models are difficult to effectively extract long-term dependency information from a large amount of data collected by sensors and don’t have sufficient learning ability.In view of the large amount of data,how to choose a reasonable model for feature extraction and long-term dependent information learning of PM_2.5 data has become the focus of current research on PM_2.5 concentration prediction.In this thesis,the research status of PM_2.5 concentration prediction is systematically analyzed,and the prediction model of PM_2.5 concentration is established by using machine learning model.The main research contents are as follows.In view of the non-stationary and nonlinear characteristics of PM_2.5 concentration,this thesis proposes an IDCNN-BGRLSTM model based on iterated dilated convolutional neural network（IDCNN）and bidirectional gate recurrent long short-term memory（BGRLSTM）network to predict PM_2.5 concentration.Firstly,the local trend features of time series are extracted by using the dilated convolution in an iterative manner.Then,the BGRLSTM model is constructed by stacking the bidirectional long short-term memory network and bidirectional gate recurrent unit network to learn the long-term dependency information and obtain the final prediction results.Relevant simulation experiments show that IDCNN-BGRLSTM model has good feature extraction effect and long-term dependency information learning ability when processing time series data.Since rainfall and snowfall data were not used for prediction,and rainfall and snowfall would have an impact on PM_2.5 concentration,this thesis further utilizes rainfall and snowfall data to establish a calibration model based on stacking.Firstly,appropriate historical data are selected by correlation coefficient.Then,the historical PM_2.5concentration,rainfall,snowfall,and the prediction results of IDCNN-BGRLSTM model were combined as the input of the base model.Finally,stacking is used to integrate the prediction results of the base learner to calibrate the prediction results of IDCNN-BGRLSTM,so as to further improve the prediction accuracy of the model.The simulation results show that the rainfall and snowfall data have a certain influence on PM_2.5concentration,and the calibration of IDCNN-BGRLSTM prediction results with stacking combined with rainfall and snowfall data can improve the prediction accuracy of PM_2.5concentration to a certain extent.