<p>Resource allocation in Software-Defined Vehicular Networks (SDVNs) remains a major challenge due to the dynamic and heterogeneous nature of vehicular environments. Traditional methods often fall short in accurately predicting resource demands under fluctuating traffic conditions, particularly when contextual factors are overlooked. In this paper, we propose a novel Federated Adaptive Attention-based Bidirectional Long Short-Term Memory (FABLSTM) model to enhance resource prediction and task allocation in SDVNs. Each local controller employs an ABLSTM network to estimate resource needs within its subnetwork, while the federated learning structure enables decentralized training and knowledge aggregation across the network. By integrating contextual variables and long-term traffic patterns through an adaptive attention mechanism, the model achieves higher adaptability to real-world dynamics. Simulation experiments based on urban traffic data from Tehran show that the FABLSTM approach outperforms baseline methods, achieving a 13.64% reduction in energy consumption, a 13.72% decrease in response time, and an 18.92% increase in throughput. These results demonstrate the model’s efficiency in managing computational, memory, and bandwidth resources under varying traffic loads. The proposed framework provides a scalable and intelligent solution for optimizing resource allocation in SDVNs. Furthermore, the large-scale and real-time nature of SDVN environments, characterized by massive vehicular data streams and distributed control layers, inherently requires high-performance computing (HPC) and parallel processing capabilities. The proposed FABLSTM framework is designed to be aligned with this paradigm, ensuring scalability and timely decision-making.</p>