Abstract
Urban traffic management in Indian Tier-2 cities encounters a fundamental forecasting challenge: demand at any given signalised intersection is not solely driven by local conditions but by festival-induced flows that propagate along inter-city corridors over multiple days. Existing single-node time-series models and fixed-adjacency graph networks cannot represent this cross-city demand propagation. This paper introduces PCU-ST-GCN, a spatio-temporal graph convolutional network specifically designed for heterogeneous Indian traffic. Node feature vectors are constructed from five IRC:106-1990 Passenger Car Unit sub-components — two-wheeler, three-wheeler, car, light commercial vehicle, and heavy goods vehicle — encoded separately to enable vehicle-class-specific propagation learning. A Cultural Event Propagation Module (CEPM) applies learnable element-wise gating conditioned on festival severity level and a fourteen-day pre-festival countdown, encoding the gradual demand build-up characteristic of major Indian festivals such as Diwali and Kumbh Mela. An adaptive adjacency matrix A* is learned end-to-end from node embeddings, recovering functional cross-city dependencies from intersection count data alone. The model is trained and tested on 69,121 fifteen-minute observations at twelve signalised intersections across the Kanpur-Lucknow-Prayagraj National Highway 19 corridor collected from June to November 2025. PCU-ST-GCN achieves Mean Absolute Percentage Error of 7.2% at the t+15 minute horizon, representing a 25% relative improvement over standard ST-GCN, with R²=0.92. CEPM reduces Kumbh Mela prediction error at Prayagraj nodes by an average of 4.1 percentage points, with the largest gain of 5.8 percentage points at Zero Road — the primary Mela access route. Ablation study confirms that PCU decomposition, adaptive adjacency, and CEPM each contribute independently to overall accuracy improvement.