Dynamic Graph Representation Learning for Spatiotemporal Forecasting under High Data Volatility

Abstract

Spatiotemporal forecasting plays a critical role in managing large-scale physical, economic, and digital infrastructures. Applications range from urban traffic management and financial order book modeling to smart grid load balancing and climate monitoring. Modern spatiotemporal architectures—primarily Spatiotemporal Graph Neural Networks (ST-GNNs)—rely heavily on spatial topology graphs. These graphs are typically constructed using physical proximity, structural connectivity, or long-term historical correlation. However, in non-stationary real-world environments, systems experience frequent high-volatility events, such as traffic accidents, sudden financial market crashes, localized weather anomalies, or unexpected power grid surges. Under these volatile conditions, static spatial assumptions fail. Spatial dependencies undergo sudden non-linear shifts, causing traditional static and parameter-bound adaptive graph neural networks to experience significant performance degradation. This paper presents the Volatile Dynamic Graph Network (VDG-Net), an adaptive end-to-end spatiotemporal representation learning framework designed for non-stationary environments under high data volatility. VDG-Net combines a Dynamic Topology Generation Module (DTGM), a Decoupled Spatiotemporal Gating Layer, and a Distributionally Robust Loss (DR-Loss) with adaptive variance regularization. Evaluated across three real-world volatile datasets (METR-LA-VOL, PEMS-BAY-VOL, and NE-GRID), VDG-Net achieves up to a 14.2% reduction in Mean Absolute Error (MAE) and a 16.8% reduction in Root Mean Squared Error (RMSE) over competitive baselines.