Urban Air Quality Prediction Based On Metalearning Under Incomplete Data

In recent years,with the continuous improvement of people’s quality of life,a high-quality living environment has become one of the factors that people pay attention to.Achieving accurate prediction of air quality is of great significance for local governments and residents to respond in a timely manner.As air quality issues are affected by a variety of complex factors,there is still a long way to go before prediction methods can be applied to actual projects.Existing research has the following problems: First,the distribution of urban air quality monitoring stations is uneven,and the existing methods are insufficient in capturing the temporal and spatial correlation between them;second,the training of existing prediction models is often based on a large number of and complete data sets.In fact,some urban monitoring stations have equipment failures or incomplete data sets,resulting in no complete data samples for model training;third,existing methods of solving data missing problems in data pretreatment phases and it is difficult to deal with a lot of data samples.In order to cope with these challenges,this article has done the following research:(1)In order to capture the spatial correlation between urban monitoring stations,this paper constructs a graph structure to represent the dependence between monitoring stations based on the geographic coordinates of the monitoring stations and generates input feature vectors based on the graph structure.Finally,experiments are conducted to verify the influence of the graph structure on the final prediction results.(2)A spatiotemporal prediction model(GAT-LSTM)based on graph attention and long and short-term memory network is proposed.The graph attention network is used to extract the correlation coefficients between monitoring stations.The graph attention network can use the attention mechanism according to the neighboring nodes.The features are assigned weights.Next,the feature vector fused with spatial information is input into the two-layer long and shortterm memory network to extract temporal correlation.(3)In response to the problem of air quality prediction,this paper uses a variety of data fill methods in the data pretreatment phase,and further proposes a time and space prediction model(Meta GAT-LSTM)based on a meta-learning,which enables GAT-LSTM model.From multiple cities to migrate knowledge,initialize the model,and use the air quality prediction of the target city to address the problem of not fully brought.(4)In this paper,there is a large number of comparative experiments in the real data set.First,the ability to predict the length of air quality in the future is evaluated by comparing the GAT-LSTM model spatial feature capture capability.Secondly,by modifying the parameters,different urban monitoring site spatial graphics structures are generated,and the forecast results under different map structures are discussed,thereby finding an optimal parameter value.Finally,urban air quality prediction in the case of data lacking,combined with data filling method,using the meta-LSTM model from multiple urban migration knowledge,initialize the GAT-LSTM model,predict the air quality of the target city on this basis.