Urban Air Quality Prediction Based On Recurrent Neural Network

In recent years,air pollution dominated by PM2.5 particulate pollutants has caused smog weather throughout China,seriously jeopardizing people's living environment and physical and mental health.According to the forecast results,the environmental protection department can provide scientific protection for heavy pollution weather and control heavy polluting enterprises.The public can take effective protection measures based on the warning information issued.Therefore,the prediction of PM2.5 concentration has important strategic significance and value to the environmental protection department and society.The formation mechanism of PM_2.5 particulate pollutants is complex,with many genesis and contains time dimension information and complex nonlinear relationships.Based on the existing PM_2.5 concentration prediction model research,on the one hand,it fails to solve the problem of input dimension disaster and uncorrelated factor interference;On the other hand,there is no effective exploration of the dependencies of time series information and the prediction accuracy caused by network performance is not high.In addition,the large sample data set consisting of air pollutants and meteorological factors has missing values of time series.Therefore,based on the above problems in the PM_2.5.5 concentration prediction model,a prediction model based on genetic algorithm and gated loop unit is established.The specific research contents are as follows:?1?Based on the air pollution data and meteorological factors of Mianyang City as the large sample data set,this paper designs a weighted sum based on the logic values of“transient value”and“steady value”for the missing values in the large sample data set.Missing value filling algorithm;?2?Aiming at the problem of dimensional disaster and uncorrelated factor interference,the search based on genetic algorithm seeks the optimal individual,that is,the input feature with high correlation;?3?Aiming at the problem of the dependence of time series information before and after the low-precision problem,the model based on the depth neural network of the gated recurrent unit is studied.Through the deep training and learning of the large sample data set,the concentration and cause of PM_2.5 are The complex nonlinear relationship and timing information extraction have better extraction and optimization capabilities;?4?For the problem of parallel processing of large sample dat a for the accelerated processing of the model,the Tensorflow-GPU platform is used to train the model and the super-parameter tuning of the designed genetic algorithm and the gated loop unit neural network prediction model?GA-GRU?;?5?For the predicted model performance problem,compare the prediction results and performance with a single GRU network model and LSTM network model.