A privacy-preserving personalized federated reinforcement learning framework with heterogeneous EV willingness for V2G scheduling

Electric vehicles (EVs) are becoming increasingly important flexible resources for grid regulation. However, the practical deployment of vehicle-to-grid (V2G) scheduling is critically constrained by a fundamental conflict between heterogeneous consumer behaviors needed for scheduling and the protection of user data privacy. Conventional centralized optimization either violates privacy or fails to capture heterogeneous charging preferences, leading to low participation willingness and unsustainable grid services. To address these challenges, this study proposes a federated reinforcement learning algorithm that enables cross-user collaborative optimization without exposing raw user data. User V2G participation willingness (VPW) is explicitly incorporated into the scheduling process through a discharging bidding mechanism, allowing heterogeneous charging preferences (e.g., arrival and departure times) to be systematically captured and enabling differentiated allocation of discharging resources. Consequently, this integration resolves the heterogeneity-privacy dilemma by converting disparate user preferences into self-organized operational roles. The approach achieves a charging cost of 0.586 CNY/kWh, representing a 56% reduction compared with conventional uncoordinated charging. Crucially, the mechanism drives a natural division of labor: high-willingness users organically assume peak-shaving duties (earning 7.84 CNY/vehicle/day), while low-willingness users default to valley-filling. Mutual information analysis confirms that this personalized, role-based scheduling is achieved with negligible privacy leakage (MI < 0.04). The VPW-based framework enables scalable and sustained participation of heterogeneous EV users under privacy constraints while preserving user autonomy and enhancing controllability of aggregated EV resources at the urban scale.