Study On The Air Quality Prediction Method Based On Spatiotemporal Attention Mechanism

In recent years,due to the industrial waste gas,automotive exhaust,and population gathering,the air quality in many areas of our country has become increasingly severe.The continuously deteriorating air quality seriously affects urban development and human health.Many cities publish real-time air quality indexes regularly by building air quality monitoring stations.However,it is not enough to know the current air quality.The society urgently needs a scientific and accurate forecasting method to provide early warning and decision support for the public and the government.On this basis,this thesis studies the air quality prediction method.First of all,there are a large number of missing values in air quality data,which will increase the training difficulty of prediction model.To solve this problem,a missing value filling method of generative adversarial networks(GAN)based on time interval is proposed.The method uses GAN to generate new complete data that conforms to the distribution of the original data set.For each missing data,the method optimizes the input random vector of the generator by introducing reconstruction loss function,space loss function and discriminative loss function,so that the generated data is most similar to the original data.The missing part of the original data is filled with the generated data.Secondly,the filled data is used for air quality prediction.The factors affecting air quality are diverse,such as the dynamic temporal correlation between different time slots,the dynamic spatial correlation between different regions,and the nonsequential influencing factors such as point of interest(POI)and road network.However,the existing prediction methods are difficult to consider the above-mentioned multiple influencing factors at the same time.To solve this problem,an air quality prediction model based on spatiotemporal attention mechanism is proposed.The model uses an encoder-decoder structure to model spatiotemporal features.A spatial attention mechanism is introduced in the encoder to capture the relative influence of different regions on the prediction area.A temporal attention mechanism is introduced in the decoder to select relatively important historical time information.Then,for nonsequential data such as POI and road network,the existing feature extraction methods tend to ignore the hierarchical relationship between different categories of POIoad network.To solve the problem,the model proposes to use the LINE method and convolutional neural network to learn the feature representation of nonsequential data.And it is used as auxiliary information to participate in air quality prediction,which effectively solves the problem of nonsequential data feature extraction and improves prediction accuracy.Finally,in order to enable the public and the government to intuitively understand real-time and future air quality changes,this thesis designs a visualization system for displaying air quality conditions.In this system,users can learn about real-time pollutant concentration information and six-hour AQI prediction results in various regions by clicking the label in the map,which can provide scientific basis for relevant departments to prevent and control air pollution,such as when to spray water to reduce dust and where to need traffic control.The thesis has 37 figures,13 tables and 80 references.