Spatiotemporal Traffic Speed Forecasting via Conditional Flow-Based Generative Deep Learning With Dynamic Graph Pooling

Abstract

Traffic speed forecasting is a crucial problem with significant real-world applications, including optimizingtraffic flow and reducing congestion. Current data-driven approaches struggle with high sample complexity, overfitting, and limited long-range temporal dependencies. To overcome these challenges, we introduce a Conditional Flow-based Generative (CFG) model with dynamic Graph Feature Extraction (GFE) and pooling for spatiotemporal traffic speed forecasting. First, traffic speed measurements are modeled as a dynamic graph, where nodes represent sensor readings and edges encode spatial correlations. A GFE network captures sparse spatial features of the dynamic graph through a hierarchical graph learning architecture, incorporating a graph pooling layer to reduce feature dimensions and a structure learning mechanism for refined graph topologies. These extracted features are fed into a peephole LSTM to capture both short- and long-term spatiotemporal dependencies. Finally, the CFG model predicts future traffic speeds of the sensors by learning the underlying data distribution, improving forecast accuracy by capturing complex patterns, and handling variability in traffic data through generative feature extraction. The proposed framework is tested using three real-world traffic datasets, PEMS-BAY, METR-LA, and PEMS07, and has demonstrated outstanding performance compared to state-of-the-art approaches. This exceptional performance is due to its ability to enhance generalization and prediction accuracy while mitigating overfitting, offering a significant advancement over recent traffic forecasting models.