GTPAN: A GRU–TPA–KAN Integrated Temporal-Pattern-Aware Neural Network for Shared Bicycle Deployment Forecasting and Dynamic Pricing Optimization

This paper develops GTPAN, an IoT-enabled temporal-pattern-aware neural framework for real-time shared mobility forecasting and dynamic pricing optimization. Within the Internet of Things (IoT) ecosystem of smart urban transportation, massive spatiotemporal data streams are continuously generated from bicycles, docking stations, and user mobile sensors. Traditional pricing and deployment schemes are often static and fail to adapt to this rapidly evolving IoT environment, leading to over-deployment and revenue loss. The proposed GTPAN framework integrates a Gated Recurrent Unit (GRU) for temporal encoding, a Temporal Pattern Attention (TPA) mechanism for adaptive feature weighting, and a Kolmogorov–Arnold Network (KAN) for nonlinear mapping, jointly enabling accurate demand prediction under complex IoT data dependencies. Furthermore, an improved domain-constrained simulated annealing algorithm is incorporated to achieve real-time dynamic pricing with low computational latency, making the model suitable for edge-intelligent deployment. Experimental results on real-world IoT-based shared bicycle datasets demonstrate that GTPAN improves mean prediction accuracy by 5.74% and revenue by 45% compared with state-of-the-art methods. The proposed framework provides an interpretable and scalable solution for IoT-driven smart mobility management, contributing to sustainable and efficient urban transportation systems.