Uncrewed aerial vehicles (UAVs) have made a substantial contribution to vehicle communications in recent times, and they provide viable ways to improve connection in contemporary transportation networks. However, maintaining consistent signal coverage, the limited computation capacity of UAVs, and getting past obstructions to maintain direct communication with vehicles is still tedious. To address the same, In this paper, an edge-enabled digital twin (DT) of UAV with an intelligent reflecting surface (IRS)-aided vehicular network is investigated. We specifically concentrate on the issue of minimizing the net energy consumption of the system in task-offloading while simultaneously optimizing IRS phase-shift, power allocation and task-offloading parameters through the use of DT architecture. We first describe the specified non-convex optimization issue as a Markov decision process (MDP) to address it. Eventually, we propose a hybrid federated learning (HFL) algorithm that aims to maximize energy efficiency (EE) by optimising related parameters. This method also enhances the system’s overall performance by lowering energy consumption and using the combined experiences of several agents. Compared to the benchmark schemes, HFL proves to be 20.5% and 47.6% more efficient than MAD2PG and DQN respectively. Simulation results affirm that the suggested method outperforms the benchmark techniques in terms of EE and learning accuracy.