The dynamic nature of traffic, high mobility and varying topology are major challenges to Urban Vehicular Ad Hoc Networks (VANETs), especially in densely populated urban areas. The methods of traditional routing and congestion control are not able to adjust and this leads to increased latency, decreased throughput and frequent breakdowns. In this paper, a new hybrid deep learning architecture that combines Graph Convolutional Networks (GCN) to extract spatial features and Long Short-Term Memory (LSTM) networks to model time is proposed with the aim of offering adaptive routing and other proactive congestion control capabilities in urban VANETs. The model applies the real time traffic information of sensors and vehicular communication systems in urban areas to forecast traffic congestion, model vehicle density and route optimization. In experimental results, the proposed GCN-LSTM model is significantly more efficient in routing and also in predicting results, compared to the baseline models. In particular, it recorded a Mean Absolute Error of 0.02, Root Mean Squared Error (RMSE) of 0.07, routing latency of 38.13 ms and Packet Delivery Ratio of 95, which was much better than GCN-only (MAE: 4.65, PDR: 90.3) and CNN-LSTM (MAE: 0.14, PDR: 88.1). The hybrid model also aids the real-time traffic processing through the edge-cloud cooperation, reducing the inference latency and scaling to city regions. Such methods as geospatial embedding and temporal batching enable the model to represent the intricate flow of traffic in high-resolution. The architecture is also modular and allows simple incorporation of both emergency vehicle prioritization and location-based rerouting and is scalable to operate on resource-constrained edge nodes, meaning it can be used in practice in intelligent city infrastructure. Altogether, the developed model offers a high-performing and scalable adaptive routing and congestion control solution to smart urban transportation systems.
Mohammed Maray (Fri,) studied this question.