Typical Air Pollutant Analysis And Prediction Based On Machine Learning

Fog and Haze are two types of extreme weather,often accompanied by low visibility or high pollution,which have tremendous impact on normal transportation and good health,as a result,it is essential to be able to forecast and warn of fog and haze in a timely manner.The main traditional methods for fog and haze research are real-time monitoring using radar and related numerical analysis methods,which have high labour costs and complex calculation processes.Machine learning,as a data mining technique,is widely used in various fields such as artificial intelligence,which provide new methods and ideas for fog and haze prediction research.However,most of the existing research work focuses on the determination of whether there is fog and the current state of haze,and less involves the classification prediction of short-time proximity fog levels and multi-step time series prediction analysis of haze.Based on the above,this thesis carries out research work on short-time proximity fog level classification prediction,first occurrence haze regression analysis and time-series haze multi-step prediction,which can achieve accurate prediction of fog and haze and can effectively compensate for the lack and deficiency of existing work.This thesis uses fog and haze data provided by meteorological centres and open data platforms for pre-processing.For fog,the class labeling is mainly determined by the visibility threshold.Because of more complex haze,this thesis focuses on PM2.5 pollutants,which are more common in haze.Based on the visibility dataset and pollutant dataset constructed from the above data,the pre-processing of the dataset is achieved through the construction of features in accordance with meteorological principles and the coding of character features in meteorological data.After classifying the fog visibility dataset according to meteorological principles,it is found that there are several categories of unbalanced data,and the unbalanced categories were enhanced utilizing the Borderline-SMOTE data expansion method,which is more sensitive to edge data.Furthermore,as Light GBM in gradient boosting decision tree has the advantages of high prediction accuracy and low training cost,a short-time proximity fog classification prediction model based on Light GBM algorithm is investigated and optimised in this thesis.A new first-occurrence pollution dataset is formed by screening the haze pollutant dataset with the first-occurrence screening principle,and a Light GBM-based first-occurrence haze regression analysis model is developed.The model is also optimised by using Bayesian optimisation with a priori model and set objective function for the complex parameters of Light GBM.A multi-input,multi-output BIGRU multi-step prediction model is developed for the time series prediction of the haze pollutant PM2.5.At the same time,the multi-step prediction results of the same moment are calculated by means of weight summation,and an Adam optimizer with adaptive updating of the parameter learning rate is used on the basis of this model.Finally,through comparative experiments,the effectiveness of the data processing methods,analytical prediction models and parameter optimization algorithms studied in this thesis is verified,and the short-time proximity fog level classification prediction,the analysis of the first occurrence of haze pollutants and the time-series prediction of haze pollutants can be better achieved.