<p>The automotive cloud systems have now become extremely dynamic with on-board units, highway infrastructure, and remote cloud servers all contributing to supporting intelligent transportation services that are latency sensitive. Nevertheless, stringent quality-of-service demands, small on-board energy budgets, and intermittent connectivity render traditional federated learning aggregation policies inefficient to execute predictive networking tasks like link quality prediction, handover prediction, and bandwidth provisioning. This paper suggests an Energy-Optimized Federated Aggregation architecture of Predictive Networking in Vehicular Cloud Architectures (EOFA-PNVC) integrating client selection, gradient compression, and an energy-aware weighting scheme with a forecasting head that handles short-horizon state prediction of networks. It uses a two-level orchestration mechanism to divide aggregation between roadside edge aggregators and a cloud meta-aggregator, a policy of residual-energy-conscious participation to throttle the contributions of constrained vehicles before convergence halts, and a convergence-of-vehicles policy. The framework is tested using a vehicular simulation testbed with an emulated 5G-V2X channel model and mixed with SUMO traffic traces and compared with five more recent federated learning baselines. Findings show mean communication energy per round decreases by 31.4% and convergence round decreases by 22.7% and an increase in network state prediction accuracy by 8.6% with equal model utility. Additional fairness is also improved in the suggested scheme among heterogeneous vehicles with diverse battery profiles indicating that a combined optimization of energy, predictive performance, and aggregation performance can be obtained without compromising on privacy assurances.</p>