Urban PM2.5 Forecasting Model Based On Support Vector Regression And LSTM

Urban PM2.5 concentration as an important air quality prediction index,the accurate prediction of this value can guide people’s production and life.At present,BP neural network and support vector regression are mainly used to predict the value,but these methods fail to fully mine the influence of environment,weather,time and other factors on PM2.5 concentration change,and the prediction accuracy is relatively insufficient.To solve the above problems,a prediction model combining support vector regression(SVR)with long-term memory(LSTM)is proposed: Morlet wavelet and polynomial mixed kernel functions are selected as the kernel functions of SVR,and the parameters are optimized by genetic algorithm(GA)and improved particle swarm optimization(IPSO).The improved model prediction results are combined with the prediction results using LSTM The final prediction result is obtained by nonlinear superposition of composite function.The main contents of this paper are as follows:(1)Aiming at the problem of incomplete orthogonal basis of RBF kernel function in traditional SVR,and the relatively insufficient result in some complex scene prediction,a hybrid kernel function SVR prediction method based on the fusion of Morlet wavelet and polynomial kernel according to the proportion of 3:2 is proposed.In the prediction effect,it retains the overall prediction advantage of polynomial kernel function,and embodies the localization and multi-resolution prediction characteristics of Morlet wavelet kernel.The results show that the SVR based on the mixed kernel function of Morlet wavelet and polynomial is more accurate than other kernel SVR.(2)To solve the problem that unknown parameters affect the model prediction in SVR,the penalty factor C and parameter sigma are optimized by using genetic algorithm.Considering that the prediction effect of the model optimized by the algorithm is still insufficient,we use the improved particle swarm support vector regression(IPSO-SVR)model to experiment again.The results show that the model based on IPSO-SVR is simple and efficient,and the accuracy is further improved compared with the model prediction results before optimization.(3)A prediction model based on support vector regression and LSTM is proposed.Because the LSTM of multivariate time series can fully mine the influence of various independent variables on the target variables under the time series,support vector regression has the advantages of less model parameters and higher prediction accuracy in the small-scale sample prediction problem,and the final prediction value is generated by the nonlinear superposition of the two parts of the results by combining functions.(4)SVR-LSTM model is applied to urban PM2.5 prediction.Different methods or models are compared with the same real meteorological pollution monitoring data set of a city.The results show that SVR-LSTM model can reflect the actual change of PM2.5 better than ARMA time series and BP neural network,and reduce 0.0244 and 0.0082 respectively in the mean square error of prediction compared with single LSTM and IPSO-SVR.It is an effective model for urban PM2.5 concentration prediction.