Research And Application On Multi-model Ensemble Methods For Forecast Of Air Quality Based On Neural Network

With the continuous development of meteorological science and technology,air quality forecast technologies have also been continuously improved.The numerical prediction methods of air quality mainly simulate the operation flow of atmospheric by the meteorological models and consider the interaction of various factors in the atmosphere in order to predict the pollutant concentration,so as to forecast the air quality.Because the uncertainty of the initial conditions of numerical forecast,the error of model itself and the chaotic characteristic of atmosphere,the results have larger prediction error.Multi-model ensemble technologies are methods to solve the uncertainty problem in numerical forecast.It can synthesize the advantage of each model and reduce the forecast error of model.The neural network ensemble methods have been applied widely to precipitation and temperature forecast,while there are few research on the ensemble models for air quality forecast.This paper proposes multi-model ensemble methods for forecast of air quality based on neural network.The applicability of each neural network ensemble model to each pollutant concentration are compared to find the most applicable ensemble models for air quality forecast.And the models are improved in their own learning algorithms.Then the improved particle swarm algorithm is used to optimize the models.The main works are as follows.Firstly,the forecasting capabilities of three meteorological models in the key cities of the Beijing-Tianjin-Hebei are evaluated,including CUACE,BREMPS and WRF-Chem.The conventional methods such as arithmetic average ensemble,weight ensemble and multiple linear regression ensemble are studied.The experiment shows that the conventional methods are obviously inadequate.Secondly,the ensemble models based on BP neural network are established,the length of training samples and parameters of BP neural network are determined by sensitivity experiments.The experimental results show that this method can effectively improve the effect of model forecast and increase the accuracy of air quality forecast compared with the single business models.Thirdly,in order to find a neural network ensemble method with higher prediction accuracy,the applicability of RBF neural network,Elman neural network,wavelet neural network and T-S fuzzy neural network ensemble models in multi-model ensemble forecast of air quality are respectively studied.Based on sensitivity experiments,the best parameters for each type of neural network model to predict PM2.5,PM10,CO,NO2,O3,and SO2 are found and fixed.Then the results of six kinds of pollutant concentrations forecasted by different types of neural network ensemble models are compared with that forecasted by BP neural network model.The experimental results show that wavelet neural network ensemble method is more suitable for multi-model ensemble forecast of air quality.Fourthly,an improved method about wavelet neural network with the additional momentum factor is proposed in this paper to solve the slow training speed of wavelet neural network to meet the requirement of business timeliness.The experimental results show that the improved wavelet neural network ensemble method can improve the efficiency of network learning,and the prediction effect is better than the original method.Fifthly,for the problem that the training of wavelet neural network is easy to fall into local optimum,the particle swarm algorithm is studied.The nonlinear inertia weight particle swarm optimization improved algorithm with mutation is proposed and then the improved algorithm is used to adaptively determine the connection weights and parameters of the improved wavelet neural network.The experimental results show that the combination of improved particle swarm optimization and improved wavelet neural network can effectively improve the prediction accuracy and reduce the prediction error of the original model.This paper proposes multi-model ensemble technologies based on neural network,which enrich the air quality ensemble forecast methods.Through a large number of comparative experiments,it is proved that the multi-model ensemble methods of this paper can effectively improve the accuracy of pollutant concentration forecast,reduce the forecast error,and provide scientific reference for government and people.