ABSTRACT

Accurate short-term traffic flow forecasting at urban intersections underpins adaptive signal control, congestion mitigation, and incident response. However, intersection-level flows are highly non-stationary due to signal timing, turn-movement interactions, and exogenous factors (weather, events). We propose a deployment-minded architecture that combines spatiotemporal convolutions over an intersection graph with sequence-level LSTMs to capture long-range dynamics. Node features (per-approach flow/occupancy, phase state, queue length proxies) are stacked into windowed tensors; a graph-aware CNN extracts localized spatial patterns, while a stacked LSTM models temporal evolution. We evaluate early, mid-attention, and late fusion of exogenous covariates (weather, calendar, incidents). On two urban datasets constructed from loop/video detectors and signal logs (?150–220 intersections), the hybrid reduces MAE by 9–17% and RMSE by 7–15% over LSTM/ConvLSTM and classical baselines across 5/15/30-minute horizons, and maintains stable performance during peak-period regime shifts. Ablations show mid-level attention fusion is most robust to missing sensors and unplanned phase splits; adaptive quantile loss improves calibration of high-quantile (p90) travel demand. A lightweight variant (depthwise ST-CNN + single-layer LSTM) sustains sub-30 ms inference per horizon on an edge CPU, enabling cabinet-level deployment. We discuss integration with SPaT data and RL signal control, fairness across corridors, and guardrails for drift monitoring. Results suggest ST-CNN + LSTM hybrids provide a strong accuracy–latency trade-off for real-time intersection forecasting in smart city operations.