Intelligent Spatiotemporal Traffic Optimization Using Federated Model Based Reinforcement Learning for Smart Cities

Traffic congestion remains a critical challenge in urban mobility, increasing travel time, fuel consumption, and environmental impact. Traditional static timing traffic signal systems lack adaptability to dynamic traffic conditions. To address this, we propose DMFLADPPRL, a novel traffic signal control framework that combines three core innovations: (i) a multilevel deep federated learning architecture that enables decentralized training across edge, fog and cloud layers; (ii) AutoDeepCABAC, a custom autoencoder-based entropy compression technique that significantly reduces communication overhead; and (iii) a modified Proximal Policy Optimization (PPO) algorithm tailored for adaptive traffic control. Unlike previous approaches, our framework tightly integrates these components to achieve privacy-preserving, energy-efficient, and real-time decision-making. AutoDeepCABAC enables seamless deployment in consumer electronics based on the Internet of Things (IoT), such as smart cameras and roadside sensors, while the hierarchical FL design ensures scalability in large urban networks. Evaluations of benchmark datasets (SIND, PeMS, Bosch Small Traffic Lights) under IID and non-IID conditions show that DMFLADPPRL outperforms baseline methods in reducing vehicle waiting time, improving throughput, and minimizing congestion. This work demonstrates the potential of intelligent and cooperative learning systems to power the next generation of smart city traffic infrastructure.