The integration of intelligent transportation systems with next-generation consumer electronics, ranging from electric vehicles to autonomous navigation tools, has created an urgent need for reliable, long-range urban mobility prediction. However, the accuracy of existing models degrades over extended horizons due to compounding errors and an inability to adapt to dynamic, evolving traffic patterns. To address this, we propose the Generative Spatio-Temporal Graph Network (G-STGN), a novel deep learning framework that synergistically integrates learnable temporal decomposition with conditional generative adversarial networks (cGAN) for stable, long-horizon forecasting. Our architecture first purifies complex traffic signals using a learnable multi-scale temporal decomposition mechanism, which adaptively separates flow data into interpretable trend, periodic, and residual components. A dynamic graph spatio-temporal encoder then models each component with tailored, evolving graph structures to capture nuanced network dependences. The core innovation is a generative prediction enhancement framework, where a cGAN learns the macroscopic distribution of future states. This provides a forward-looking guidance signal to proactively correct the step-by-step prediction process, effectively suppressing error propagation. Extensive experiments on major highway and urban mobility datasets confirm that G-STGN significantly outperforms five state-of-the-art baseline models. It achieves superior prediction accuracy, evidenced by substantially lower Mean Absolute Error and Root Mean Square Error, alongside faster and more stable convergence. This demonstrates the framework’s strong potential for reliable deployment in time-critical consumer electronics applications, paving the way for more efficient routing, improved energy management, and safer autonomous navigation systems.