Subway passenger flow prediction based on domain-adaptive dynamic graph convolution
In response to the challenges in passenger flow prediction stemming from limited data,which led to issues of high variance and poor generalization during model training,we propose a Domain Adaptive Dynamic Graph Convolutional Network(GCN-DANN).This approach involves constructing a node network topology for metro lines and utilizing dynamic graph convolutional networks to extract correlated features,such as flow between adjacent stations,traffic load of stations belonging to the same line,and flow propagation between different lines.Additionally,we employ transfer learning to adaptively align features between the source and target domains,mitigating the performance degradation caused by inconsistent data distributions.Finally,through a fully connected layer,we integrate features from both domains to address deficiencies,such as high variance and poor generalization,observed during the training process.The model is trained on the Shenzhen metro dataset and tested on both the full and 20%sample datasets of the Hangzhou metro.Experimental results demonstrate that on the 20%sample dataset,compared to classical forecasting networks,the GCN-DANN network achieves an average reduction of 5.34%in MAE,6.07%in RMSE,and 2.97%in MAPE.On the full sample dataset,GCN-DANN exhibits decreases of 2.76%in MAE,1.77%in RMSER,and 3.5%in MAPE,based on the 20%sample.Compared to other classical networks,it experiences the smallest decrease in performance.This research successfully addresses the challenge of poor prediction performance in practical applications due to sparse data availability.
NETL
NSTL
万方数据
