Spatial-Temporal Graph Convolution Neural Network For Fine-grained Air Quality

With the development of industry and the agglomeration of population,air pollution has dramatically influenced the ecological environment and human health.Consequently,accurate and timely prediction of air condition seems significant in terms of facilitating policy formulation and minimizing potential losses.However,air quality forecasting is a challenging task due to dynamic spatial-teoporal dependency.Motivated by those,this paper mainly focuses on the air quality prediction.The main innovations and work of this dissertation are as follows:Firstly,to capture the spatial dependency,the air quality monitoring stations are regarded as the nodes and leveraged to establish a graph network.Furthermore,different from the previous approaches on the graph neural network,the paper proposes a dynamic graph neural network(DyGraphNN),which leverages the station-level attention machenism to learn the importance of a station and its neighbors changing with the environment and time.The dispersion of air pollutants between different locations is simulated as a bidirectional random walk on the dynamic directed graph.Sencondly,for the precise long-term prediction,this paper combines the deep recurrent neural network with the dynamic graph convolution neural network to predicte the air quality.Meanwhile,the paper incorporates external factors for the affect of the air quality and adopts the scheduled sampling technique during training,which improve the accuracy of long-time prediction.Thirdly,this model is evaluated on two real-world air quality datasets,and experimental results demonstrate that the proposed approach outperforms several baselines by a significant margin.It is worth noting that,with this approach,this algorithm received the first place prize out of 3874 teams for air quality prediction in the Knowledge Discovery and Data Mining(KDD)Cup challenge 2018.