PRoFL-IoV: A privacy-preserving and robust federated learning framework for short-term load forecasting in the internet of vehicles

The growing integration of Electric Vehicles (EVs) within the Internet of Vehicles (IoV) paradigm has made Short-Term Load Forecasting (STLF) a critical component for efficient energy management in modern power grids. Accurate STLF enables proactive grid operation, optimizes energy generation and distribution processes, and facilitates seamless vehicle-to-grid (V2G) interactions. However, traditional centralized load forecasting approaches are less effective in the IoV context due to privacy concerns and limited adaptability to the dynamic, heterogeneous, and resource-constrained nature of mobile EV clients. To address these challenges, we propose PRoFL-IoV, a unified deep-forecasting, privacy-preserving and robust federated learning framework for STLF in IoV environments. The framework employs a deep forecasting model composed of a Temporal Convolutional Network (TCN) to extract local temporal patterns and a Gated Recurrent Unit (GRU) to model long-term dependencies. A dense layer is then used to combine these layers to accurately predict temporal patterns in distributed load data. Next, to ensure data privacy and confidentiality, we incorporate Differential Privacy (DP) through Gaussian noise perturbation and a secure aggregation mechanism that masks client updates. Furthermore, to enhance resilience against adversarial behavior, we use a robust aggregation mechanism that filters out anomalous updates and mitigates the impact of non-Independent and Identically Distributed (non-IID) or poisoned data. Moreover, to improve efficiency, we introduce a delay plus energy-aware client selection mechanism that dynamically selects only those EVs capable of timely, computationally feasible, and sustainable participation in each STLF model-training round. Finally, a federated learning based algorithm is proposed that seamlessly integrates all these components into a coherent, end-to-end systematic framework. We implement and evaluate the proposed PRoFL-IoV through a hybrid environment, using the Pecan Street dataset for realistic energy usage patterns, simulating system-level behavior and network conditions with NS-3, and training the model in Python. The experimental results demonstrate that PRoFL-IoV performs better compare to benchmarks. It achieves a 7.69% reduction in term of Mean Absolute Error (MAE), 5.33% improvement in attack resilience, 8.33% lower prediction error under heterogeneity, 7.14% reduction in communication overhead, and an 11.11% improvement in energy efficiency.