Stability-Aware Dual Graph Recurrent Network for Citywide Wireless Traffic Forecasting

Accurate and stable wireless traffic forecasting is crucial for proactive resource management in emerging urban Internet of Things (IoT) systems. However, existing approaches struggle to capture the complex interplay between fixed physical topologies and latent, data-driven spatial correlations. Moreover, dynamically learning graph structures often introduces optimization instability. In this paper, we propose the Dual Graph Recurrent Network (DGRN), a forecasting model that processes input signals on a fixed physical graph while evolving hidden states on a learnable latent graph. To ensure stable learning, we theoretically prove that DGRN’s output is Lipschitz-stable with respect to perturbations in the latent graph, and we explicitly relate the stability bound to the model’s structural hyperparameters. Building on this result, we develop a bilevel optimization framework that jointly learns the latent graph and model parameters with guaranteed convergence. Extensive experiments on three real-world wireless traffic datasets demonstrate that DGRN achieves state-of-the-art forecasting accuracy and learns meaningful latent graph structures. The empirical evidence confirms DGRN’s stability and convergence, highlighting its practical potential for urban IoT forecasting systems.