Research On Temporal-Spatial Hybrid Prediction Method Of Air Quality

Burning of household fossil fuel,emission of industrial waste gas and vehicle exhaust have caused serious air pollution in recent years,which has raised widespread concerns of the whole society.Air pollutants have a bad effect on human health and social sustainability.Many cities build air quality monitoring stations and publish air quality index(AQI)to people regularly.However,it is not enough to just monitor AQI,and a reliable and steady air quality prediction method is desperately needed by society to provide decision support to governments,businesses and people.This thesis centred on air quality prediction method and proposed a temporal-spatial hybrid prediction method of air pollutant based on CGMM(TSHM-CGMM),which consists of three parts: a temporal predictor based on DWT&SVR(TP-DWT&SVR),a spatial predictor based on ELM(SP-ELM)and a prediction aggregator based on CGMM(PA-CGMM).TP-DWT&SVR first decomposes local AQI into high-frequency and low-frequency signals of different scales,and then performs nonlinear regression analysis of the high-frequency and low-frequency signals and other local dada in each scale by using support vector regression.TP-DWT&SVR solves the problem that existing temporal predictor just considers the linear features of local data.SP-ELM learns global data fast based on the high learning speed and strong generalization of extreme learning machine,which solves the problem that existing spatial predictor has a long training time and prediction time.PA-CGMM clusters the local meteorological data first,and then builds a probabilistic Gaussian Mixture Model of the aggregate prediction according to the prediction deviations of TP-DWT&SVR and SP-ELM in each meteorological class.PA-CGMM calculates the optimal weights of each Gaussian Mixture Model according to Expectation Maximization Algorithm,and aggregates the predictions of TP-DWT&SVR and SP-ELM dynamically,which solves the problem that existing prediction aggregator has a bad aggregate result because of too many uncertain parameters.The experimental results prove the reliability and high efficiency of TSHM-CGMM.PA-CGMM only aggregates the predictions of TP-DWT&SVR and SP-ELM from the aspect of meteorological class,while not considers the effect of different prediction environment in the same meteorological class and meteorological metrics on the aggregate prediction,which results the systematic forecast deviations.To solve the above problem,this thesis proposed an air quality forecast deviation correction method based on similar forecast kalman filter(FDCM-SFKF).FDCM-SFKF first searches a few most similar forecasts with the current forecast environment from the training set of the current meteorological class,and then implements kalman filter algorithm in the order of smallest similarity to the largest similarity,and finally estimate the optimal forecast deviation of PA-CGMM under the current prediction.FDCM-SFKF learns its parameters by using the corresponding training set,which ensures the valid and optimal correction on PA-CGMM's forecast deviations,and improves the prediction accuracy of TSHM-CGMM.The experimental results prove the validity of FDCM-SFKF.