Research On Ozone Concentration Prediction Based On Deep Learning And Spatio-temporal Hybrid Ensemble Method

With the rapid development of China’s industry,the environmental problems caused by air pollutants represented by PM2.5 and O3 have received widespread attention.In recent years,although the effect of PM2.5 pollution control is remarkable,the concentration of O3 is still at a high level for a long time.The prediction of O3 pollutant concentration can provide a technical basis for preventing its concentration from exceeding the standard in advance and emergency treatment of high-pollution weather.At the same time,it is also of great significance to public health and government decision-making.However,the current conventional O3 concentration prediction is more concerned with short-term prediction,and the reliability of medium and long-term prediction needs to be improved,so it is difficult to fully play the role of early warning.In addition,some data-driven O3 concentration prediction models are not interpretable enough to meet the needs of supporting relevant management decisions.Therefore,how to achieve high-quality and interpretable O3 concentration prediction is one of the difficult problems in the field of atmospheric environmental governance.In response to these difficulties,this paper improves the data-driven O3 concentration prediction model from the perspective of O3’s own characteristics and generation mechanism,from the perspectives of features,models,and interpretability,and proposes a O3 concentration mixture based on spatiotemporal feature extraction.Integrate the model to improve the overall prediction level of O3 concentration and reveal the key factors in the process of O3 concentration prediction.The specific research work and related achievements are as follows:First of all.this paper proposes an automatic data preprocessing and feature extraction method.The data preprocessing strategy is determined through data distribution,and feature selection is performed in combination with the ozone generation mechanism,correlation analysis,and feature importance analysis results,which improves the prediction effect of the model.Secondly,a mixed integration model of O3 concentration based on spatio-temporal feature extraction is proposed and the parameters are optimized.The model includes a deep learning module,a meteorological prediction module,a spatiotemporal interpretation module and an integration module.Each module uses different inputs to capture and utilize time,space and weather information in a targeted manner.The integration module organically combines all parts to obtain the final prediction results.Finally,this paper takes Hangzhou Binjiang monitoring station as the experimental object,evaluates the forecast results of the spatio-temporal hybrid integration model from multiple perspectives such as typical weather forecast results,24-hour forecast accuracy and actual IAQI forecast accuracy,and compares them with MLP,GRU,LSTM,The attentionbased Seq2Seq model is compared with the prediction results of the weather-coupled XGBoost model.The results show that the spatio-temporal hybrid ensemble model has the best performance in terms of peak value and change trend prediction under different O3 pollution weather,MAE,RMSE and R2 under different time scales,or in terms of actual IAQI prediction accuracy.Specifically,its MAE in single-hour O3 concentration prediction is 5.94 μg/m3,which is 29.4%and 15.3%lower than LSTM or Attention-based Seq2Seq models,respectively.In the 24-hour comprehensive O3 concentration prediction,its MAE is 19.35 μg/m3,which is 7.4%and 2.7%lower than the latter two models,respectively.To sum up,the mixed integration model of O3 concentration based on spatio-temporal feature extraction proposed in this paper can capture the spatio-temporal correlation in the data well.While improving the interpretability,it also effectively improves the overall prediction effect of the model.The field of concentration prediction has certain practical value and theoretical value.