A federated relational-graph multi-agent reinforcement learning framework, termed FedRG-MARL, enables decentralized electric vehicle scheduling across heterogeneous park microgrids while respecting physical transformer boundaries and local data privacy. In simulated multi-park microgrid systems connected through a shared regional transformer, the framework achieved a 76.3 percent increase in mean daily system profit compared to the constrained graph-based baseline GNN-CSAC over 150 stochastic test scenarios. The scheduling architecture allows autonomous district controllers to balance local charging station demand and vehicle-to-grid operations while ensuring aggregate grid exchanges remain within regional transformer capacity limits without exchanging raw operating trajectories.
The study examines the operational challenges arising from the rapid integration of electric vehicles, distributed photovoltaics, and battery energy storage systems into regional distribution networks. When multiple park-level microgrids share a single regional distribution transformer, autonomous local controllers face strict capacity-coupling constraints at the grid interface alongside heterogeneous local demand patterns. At the same time, operating data—such as individual driving records, device operating states, and commercial tariffs—remain commercially sensitive and subject to local governance boundaries. Conventional centralized optimization techniques like nonlinear programming and model predictive control struggle to scale under privacy barriers, while standard federated learning algorithms suffer from negative transfer under non-independent and identically distributed park conditions.
To overcome these physical and informational barriers, the authors formulated the cooperative scheduling challenge as a multi-agent constrained Markov decision process. Each park microgrid acts as an autonomous agent tasked with generating continuous control commands for its local controllable battery storage and active vehicle connections. Rather than sharing complete local state trajectories, each decentralized agent constructs a local relational graph observation capturing interactions among heterogeneous components. This graph explicitly links electric vehicle nodes, a stationary battery energy storage system node, a local charging station node, an exogenous state node, and a regional transformer feedback node, preserving relationship-specific physical and scheduling semantics.
The relational graph neural network incorporates typed nodes and relation-specific message-passing channels with shared relation transformations and park-private low-rank adapters. This shared-private architectural design separates transferable physical interaction knowledge from park-specific operational dynamics. Incoming messages across active relation channels are controlled via relation-wise normalization, which mitigates numerical scale imbalance across heterogeneous connections while allowing charging station features—such as pile occupancy and net-demand pressure—to retain aggregate demand-intensity information. Adaptive relation fusion weights active channels dynamically, while a role-preserving critic readout pools node representations by functional type, preventing sparse transformer security feedback from being obscured by high-volume vehicle data.
To guarantee operational safety without relying solely on post-hoc intervention, FedRG-MARL employs a dual-level constraint management approach. An execution-side safety mechanism enforces device-level bounds, charging station capacity, and regional transformer limits prior to physical grid interaction. Concurrently, learning-side twin cost critics estimate expected future constraint violations based on policy-requested, pre-correction exchanges. A local-regional constraint-cost decomposition penalizes both local charging station overshoots and aggregate transformer capacity exceedances through an adaptive Lagrangian multiplier update. This guides decentralized policies toward naturally feasible scheduling decisions during training, reducing reliance on external correction mechanisms.
Cross-park knowledge collaboration is conducted through a trusted third-party federated aggregation layer using a target-conditioned evaluation sequence. Participating parks periodically submit restricted evaluation packages containing only actor-side shared parameters, compatibility-evaluation networks, and small observation mini-batches, while raw trajectories, replay buffers, and optimizer states remain strictly within local domains. The trusted layer evaluates source parameter advantages against the target park’s specific reward and constraint objectives, computing normalized compatibility signals to establish personalized source weights. Candidate models undergo re-evaluation and gated loading, wherein beneficial parameters receive adaptive updates, near-neutral updates undergo probing adjustments, and unfavorable transfers are rejected.
To protect regional real-time operational coordination from exposing individual park behavior, the framework incorporates a pairwise additive-masking abstraction. Local pre-coordination exchanges and controllable capacity flexibilities are masked using canceling pairwise random values before regional aggregation. The regional operator calculates aggregate overload status, direction, and total flexibility without observing individual district contributions. If an overload occurs, proportional adjustments are distributed to each park based on aggregate capacity metrics, enabling local controllers to curtail charging, trigger discharging, or adjust photovoltaic generation using internal device limits.
The simulation environment modeled three distinct, highly heterogeneous district microgrids—residential, office, and commercial—connected to a shared distribution transformer. The districts featured different charging station ratings, battery capacities, time-of-use tariffs, and vehicle arrival models based on truncated Gaussian distributions. Photovoltaic generation was installed exclusively in the office and commercial parks with weather-conditioned scaling across sunny, cloudy, overcast, and rainy days. A simulated fleet of 20 distinct electric vehicle models captured realistic battery capacities, state-of-charge charging power taper curves, and vehicle-to-grid capabilities, ensuring the learning algorithms operated across non-identical, highly dynamic environments.
Across 150 common stochastic evaluation scenarios spanning five independent evaluation seeds, FedRG-MARL achieved a mean daily system profit of 1786.51 with a standard deviation of 152.45, corresponding to a 95 percent confidence interval between 1761.92 and 1811.11. In paired scenario comparisons, the proposed method improved mean profit by 23.41 percent over MLP-SAC, 76.28 percent over GNN-CSAC, and 79.30 percent over GNN-SAC, consistently outperforming each baseline in all 150 matched test scenarios. Simultaneously, users’ mean daily net charging expenditure remained stable at 1644.36 with a standard deviation of 81.32, demonstrating that profitability gains resulted from improved temporal load shifting, solar self-consumption, and vehicle-to-grid dispatch rather than higher user costs.
Controlled comparisons and ablation analyses confirmed the distinct contributions of each architectural component. In tests on identical learning backbones, FedRG-MARL increased overall mean daily profit by 5.72 percent compared to a non-federated variant, while conventional FedAvg and FedProx failed to provide benefits due to uniform parameter aggregation under non-IID conditions. Progressive functional ablations demonstrated that introducing shared-private relation-aware representations improved mean profit by 45.88 percent over homogeneous graph baselines, enhanced relation-integration and role-preserving critics added 14.30 percent, and personalized federation contributed an additional 5.72 percent gain. Post-training diagnostics revealed zero charging station capacity violations for FedRG-MARL, validating its ability to maintain physical feasibility under local privacy constraints.
Detailed operational profiles across representative testing days illustrate how the coordinated agents translate physical flexibility into financial gains. Under uncoordinated baseline strategies, electric vehicle charging demands caused severe localized spikes, particularly during evening residential return periods and morning office arrivals. In contrast, FedRG-MARL shifted charging schedules to exploit prevailing conditions, increasing grid electricity imports around midday when solar generation was active, transitioning toward net power export in the afternoon, and maintaining near-zero exchanges overnight. Across the individual microgrids, the framework utilized stationary storage systems as dynamic load-shifting resources rather than letting them sit idle at low energy levels. Battery units were charged between midday and afternoon windows in response to park-specific tariff signals and subsequently discharged ahead of higher-demand or higher-priced periods. Concurrently, vehicle-to-grid discharging actions were selectively dispatched to balance regional capacity constraints, while residential vehicles were primarily charged during late-night intervals.
These temporal shifts were driven by structured economic price incentives and verified through fine-grained diagnostic tests. Pricing was modeled on the Jibei grid commercial time-of-use tariff structure, featuring peak, flat, valley, and deep-valley price bands adjusted across ninety-six daily decision steps. To isolate policy learning from physical enforcement layers, researchers conducted a diagnostic evaluation over thirty stochastic scenarios comparing policy-requested actions before and after safety interventions. The unconstrained baseline GNN-SAC generated a charging station pre-correction violation rate of 0.5903 percent, with an average exceedance of 0.05358 kilowatt-hours per park-step requiring 0.10580 kilowatt-hours per park-step of local safety correction. In comparison, both GNN-CSAC and FedRG-MARL recorded zero pre-correction capacity exceedances across all evaluation steps, demonstrating that the Lagrangian cost-critic formulation successfully guided neural policies to internalize infrastructure constraints. Furthermore, regional transformer adjustments in normal scenarios were negligible, measuring between 10⁻¹⁴ and 10⁻⁸ kilowatt-hours daily due purely to numerical and aggregation rounding.
System-level profiling on an Intel Core i5-14400F processor provided concrete benchmarks for communication payloads and computational scalability. In the three-park setup, each participant uploaded a restricted evaluation package of 3.31 megabytes per round and received an updated shared backbone of 0.327680 megabytes, yielding a total logical federation traffic volume of 10.91 megabytes per cycle. For real-time grid coordination, protecting four regional state variables required twelve pairwise masks, which executed in 0.0169 milliseconds through fixed-point additive masking without noticeable transmission lag. However, the evaluation architecture highlighted clear scaling boundaries. Because every target park evaluates all peer models across target-specific data batches, the number of ordered evaluations grows quadratically with the number of participating microgrids according to N times N minus one. For three districts, the required six pairwise evaluations took 1.220 seconds per round, establishing target-conditioned compatibility scoring as the primary computational bottleneck and emphasizing the need for more scalable federated evaluation protocols in larger regional networks.
Subject of Research: Privacy-compatible electric vehicle charging coordination in coupled multi-park microgrids using federated multi-agent reinforcement learning
Article Title: A federated relational-graph multi-agent reinforcement learning framework for privacy-compatible electric vehicle scheduling in coupled multi-park microgrids
Article References: Zhang, S., Qi, H., Chen, X., & Dong, L. (2026). A federated relational-graph multi-agent reinforcement learning framework for privacy-compatible electric vehicle scheduling in coupled multi-park microgrids. Journal of King Saud University Computer and Information Sciences, 38(9), Article 996. https://doi.org/10.1007/s44443-026-01326-w
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01326-w
Keywords: electric vehicles, microgrids, multi-agent reinforcement learning, federated learning, smart charging, power distribution transformer, relational graph neural networks, battery energy storage, vehicle-to-grid, photovoltaic integration, constrained markov decision process, privacy preservation, energy management, additive masking, distributed optimization, grid capacity constraints, charging station scheduling
Cite Scienmag News
Faith Mcneil. (November 21, 2026). Federated Multi-Agent Reinforcement Learning Optimizes Electric Vehicle Scheduling Across Microgrids. Scienmag. https://scienmag.com/federated-multi-agent-reinforcement-learning-optimizes-electric-vehicle-scheduling-across-microgrids/
Faith Mcneil. "Federated Multi-Agent Reinforcement Learning Optimizes Electric Vehicle Scheduling Across Microgrids." Scienmag, 21 November 2026, https://scienmag.com/federated-multi-agent-reinforcement-learning-optimizes-electric-vehicle-scheduling-across-microgrids/. Accessed 22 November 2026.
Faith Mcneil. "Federated Multi-Agent Reinforcement Learning Optimizes Electric Vehicle Scheduling Across Microgrids." Scienmag. November 21, 2026. https://scienmag.com/federated-multi-agent-reinforcement-learning-optimizes-electric-vehicle-scheduling-across-microgrids/








