Time Scale Research On Air Quality Index Forecasting Based On Time Series Network

In recent years,many places over the country have staged the tragedy of "ten-faced haze".The fine particles in the smog will increase the risk of cardiovascular disease and lung cancer,seriously threaten human health,and efforts to solve air pollution are imminent.The Air Quality Index(AQI)is a dimensionless index that quantitatively describes the air quality status.The larger the value is,the higher the level is,indicating that the air is more polluted.Therefore,accurately predicting AQI has become one of the important research directions to improve and control air quality.At present,air quality prediction models mainly include numerical model prediction models and statistical prediction models.The numerical model requires stronger professional knowledge for data processing and parameter configuration,which increases the research cost and complexity,and lacks applicability,therefore,the statistical prediction models are widely used.However,in the related research,on the one hand,the influence of meteorological conditions on the change of pollutant concentration was not comprehensively considered,only the pollutant factor was used to predict AQI,so the prediction accuracy was not high.On the other hand,the granularity of AQI prediction is not refined enough,and only a fixed step size is used to predict the value of AQI at a certain time in the future,so that the memory relationship of the AQI value next time and the pollutant content of the time period nearest to that time not only is ignored,the relationship between the prediction model and the prediction step also is ignored.This study considers the above three shortcomings comprehensively and proposes its own research content.First,information transfer entropy is used to screen out the key meteorological factors that affect the strong changes in air quality.Then,for the purpose of improving the prediction accuracy,for different prediction durations,models with different prediction steps are established and analyzed.Taking Chengdu hourly historical environmental data and meteorological data from January 1,2018 to January 15,2019 as the research object,the Long Short Term Memory unit(LSTM)and Gated Recurrent Unit(GRU)as the basic network structure,using historical data of the first 1-48 hours including the current time to establish different prediction step models to predict AQI values in the next 1-48 hours.After experimental comparison,preliminary conclusions are drawn.For the AQI prediction of the next 1-30 hours,the T+30 model based on the LSTM time series network needs to be separately trained,and the T+48 model based on the GRU time series network for the next 31-48 hours can be used,taking the 31-48 hours AQI prediction value.Experiments show that using such a combined model is beneficial to improve the prediction accuracy.Finally,in order to further improve the performance of the prediction model,the Adam(Adaptive Moment Estimation)optimization scheme is used to tune these models.After experimental comparison,in general,when each model predicts the AQI for the next N hours,these models based on the Adam optimization method perform better than models based on the RMSprop optimization method.To draw the final conclusion,this paper proposes that prediction of the AQI values for the next 1-30 hours can use the T+30 model of the LSTM timing network based on the Adam optimization method,and for the next 31-48 hours,can use the T+48 model of GRU sequential network based on the Adam optimization method for the AQI prediction,taking the predicted value of the next 31-48 hours.Judging from the application results,it can be considered that the combined forecasting model proposed in this paper has positive significance for the AQI forecast in Chengdu.