Research On Ozone Prediction Algorithm For Multi-source Monitoring Data

In recent years,ozone（O₃）has gradually become the primary pollutant that plagues the improvement of urban air quality.Accurate and efficient ozone prediction is of great significance to the prevention,control and treatment of ozone pollution.The air quality monitoring networks provide multi-source air quality monitoring data for ozone predic-tion tasks.However,ozone prediction based on multi-source monitoring data still faces the following three challenges:First,the prediction accuracy is impacted by the outlier stations and the redundant stations;Second,the relationships between stations will be affected by meteorological conditions;The third challenge is the cross-interference be-tween different pollutants.To solve above challenges,this thesis develops an ozone pre-diction method for multi-source monitoring data from the perspectives of evolutionary learning and deep learning.The main works for this thesis are summarized as following:（1）Aiming at the interference caused by the redundant and outlier stations,an im-proved whale optimization algorithm based on transfer entropy and mutual information is proposed to select and optimize the monitoring stations.First,transfer entropy and mu-tual information are used to quantify the causality and correlation between each stations.Secondly,we propose a dynamic priority search strategy to improve the convergence ac-curacy and speed up the algorithm.Finally,the linear convergence factor is improved to an adaptive convergence factor to balance the global search ability and local search ability by describing the search progress of algorithm using the diversity of selection probabilities between each stations.From the experimental results of 12 stations,the improved whale optimization algorithm has an optimal performance under 91.67%of stations,which effectively eliminating the interference of irrelevant stations on the pre-diction.（2）Aiming at the cross-interference between pollutants and the dynamic station re-lationships caused by meteorological conditions,we propose an ozone prediction model based on the dynamic spatio-temporal attention mechanism.First,graph embedding is utilized to learn the spatial dependence between stations.Then,feature embedding is used to express the dynamic meteorological information.After that,1-dimensional convolu-tion is introduced to solve the cross-interference problem.Secondly,the station informa-tion fusion module based on dynamic spatio-temporal attention mechanism is proposed to fuse the stations information dynamically under the dynamic meteorological conditions.Finally,the convolutional gating unit is proposed to measure the encoded information temporally.Experimental results show that,compared with other baseline models,the MAE value of the proposed model is reduced by 17.62%on average,which effectively alleviates the cross-interference between pollutants and reduces the impact of meteoro-logical conditions on ozone prediction.