ST-GACN: A Lightweight Spatio Temporal Graph Attention Model With Online Adaptation for Real Time Congestion Forecasting

Accurate and low latency traffic prediction is a backbone for modern adaptive signal and digital twin platforms. Yet, most deep models are either too static to handle concept drift or too heavy for roadside hardware. We present ST-GACN, a spatio-temporal Graph Attention Convolutional Network that freezes a sparse multi-head attention backbone and fine-tunes only a 3 k-parameter read-out layer online. The architecture couples local global graph attention with dilated causal convolutions, achieving wide receptive fields while remaining memory-thrifty. Evaluated on the METR-LA benchmark (207 sensors, five-minute granularity), ST-GACN attains <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAE</i> = 0.1752 at a five-minute horizon and 0.2511 at fifteen minutes, surpassing Diffusion CNN, Graph WaveNet, and ASTGCN by 9–13%. Ablation confirms that removing either dynamic attention or online adaptation degrades accuracy by up to twenty percent. The complete forward pass, including a single EWMA weighted gradient step every thirty minutes, executes in 15<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i> on an NVIDIA A100 and 62<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i> on a Jetson Orin, comfortably within the 100<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i> control Budget for field deployment. Qualitative diagnostics reveal that the model reweights upstream sensors within one interval during incidents, and spatial error maps localize residual uncertainty to stadium ramps. One factor sensitivity experiments reveal an optimal trade-off at four heads, a one-hour history window, and a thirty-minute adaptation cadence. The open source release includes Docker and Jetson flash images, paving the way for drift-aware, edge-resident congestion forecasting and future multimodal extensions.