Multi-Graph Spatio-Temporal Convolution for Traffic Flow Prediction Focusing on Edge Derived Imbalanced Data From Highway Electronics

The proliferation of Internet of Things (IoT) and the adoption of Edge Computing paradigms have led to a substantial increase in data volume. Data imbalance is inevitable in distributed environment composed of edge nodes, due to inherent diversity in processing capabilities, network connectivity, and geographical distribution. In specific highway domain, such edge derived data makes predictive precision of crucial traffic flow hard to guarantee. First, data imbalance would deteriorate the predictive performance at certain locations. Second, various spatial dependencies existed in highway network have not been sufficiently explored. To address these challenges in highway domain, we propose a novel method for daily traffic flow prediction on imbalanced data. On the one hand, a pre-processing strategy is presented to normalize traffic flow data in long-tail distribution. The data is collected from highway electronics through edge nodes. On the other hand, a multi-graph convolution model is designed to capture spatio-temporal features in distinct physical, statistical and latent perspectives. Incorporating external characteristics like meteorological and calendar features, prediction can be achieved precisely. Through extensive experiments and case studies conducted on a Chinese provincial highway, our method demonstrates a significant improvement of predictive accuracy than baselines and outstanding effects in a practical project.