Research On Air Quality Prediction Based On Deep Spatial Feature Extraction

In recent years,with the continuous increase of haze weather,air quality has gradually become a hot topic.Long-term inhalation of haze can endanger people’s health and cause diseases such as asthma and lung cancer.Therefore,the ability to accurately predict air quality in advance is of great significance for guiding people to travel and taking protective measures.Air quality prediction is also an important branch of the current time series forecasting task.The prediction of air quality is mainly the forecast of PM2.5.Although there are many factors that affect air quality,we can only obtain a small part.Therefore,the current research can only use a small part of the collected data for modeling.This article will These data other than PM2.5 are collectively referred to as auxiliary information.Most of the existing models put PM2.5 and auxiliary information together indiscriminately to extract temporal features,but in fact,different auxiliary information has different effects on the prediction results.This undifferentiated feature extraction cannot distinguish different auxiliary information.The importance of the data,that is,the spatial feature extraction of the data is not sufficient.For example,the GRU(Gated Recurrent Unit)model used in many works inputs all data into the model for prediction without distinction.It cannot distinguish the importance of different auxiliary information for prediction,and only considers temporal feature extraction.To solve these problems,we designed an AI-GRU unit(Auxiliary InformationGRU)that can extract the importance of different auxiliary information for PM2.5prediction,and use it as an important component module that can extract deep spatial features and integrate it with deep network transition model AI-DTN(Auxiliary Information-Deep Transition Network)that designed by us,air quality data is predicted and analyzed.The specific work is as follows:1.Aiming at the characteristics of air quality prediction tasks,in order to extract the deep spatial and temporal characteristics of air quality data,this paper proposes an air quality prediction model based on auxiliary information and deep transition network(Auxiliary Information-Deep Transition Network,AI-DTN),it contains two transition networks in different directions,which extract feature information from the two time series directions,respectively,to enhance the degree of feature extraction.Each transition network is composed of our designed AI-GRU which extracts spatial features and integrates auxiliary information,and the existing Transition GRU(T-GRU)which extracts temporal features.In AI-GRU,we have designed two kinds of gates,one controls the degree of access control loop unit of auxiliary information flow,and the other controls the degree of fusion of PM2.5 and auxiliary information.This kind of gating mechanism can avoid the information fusion process.Interfere with each other.We compared AI-DTN with LSTM,GRU,DARNN,GP and VARMLP.Experiments show that AI-DTN is better than LSTM,GRU,GP and VARMLP,which will not extract spatial information well,and is significantly better than DARNN,which also focuses on extracting spatial information.In addition,we performed ablation experiments on AI-DTN by deleting different components in AI-DTN.The experimental results proved that the gate control in our proposed AI-GRU can extract the importance of different auxiliary information to PM2.5,also proved the importance of bidirectional data input in AI-DTN in extracting time features.2.Apply AI-DTN to the air quality prediction system.Realize the air quality prediction system.The system takes the Changchun City area as an example.It reads time series information from the database.The background predicts the air quality according to the prediction algorithm AI-DTN.At the same time,the prediction results are written back to the database,and finally the system reads from the database.Take the result and send it back to the front end of the system,and display it to the user in the form of a line chart.