Privacy‐Preserving Federated Multi‐Agent Reinforcement Learning for Coordinated Transmission–Distribution Dispatch

ABSTRACT

High penetrations of distributed energy resources, electric vehicles and battery storage require closer coordination between transmission system operators (TSOs) and distribution system operators (DSOs), but centralized dispatch can expose commercially sensitive operational data. This paper proposes FedMASAC, a federated multi‐agent soft actor‐critic framework for coordinated transmission–distribution dispatch. Each DSO trains a local policy from its own observations and communicates a clipped, Gaussian‐perturbed, sparsified and quantized pseudo‐gradient to a TSO‐hosted coordinator. The coordinator combines these updates using weights based on training‐set size, boundary power exchange and voltage stress, while proximal regularization limits client drift under heterogeneous operating conditions. Each communicated pseudo‐gradient is protected by a calibrated Gaussian mechanism under the stated honest‐but‐curious threat model; cumulative privacy accounting is discussed separately from the empirical attack evaluation. Experiments on a coupled IEEE 118‐bus/33‐bus system with 10 DSO agents compare the method with independent multi‐agent learning and raw‐data federated‐learning baselines. In a reproducible reduced‐scale reproduction (10 DSOs, 100 federated rounds and five seeds), FedMASAC attains a dispatch cost above a perfect‐foresight SOCP benchmark with a voltage‐violation rate, versus for independent MASAC and for raw‐data federated learning; under non‐IID feeder heterogeneity the corresponding figures are and . Top‐ sparsification with 8‐bit quantization reduces upload volume by about 93%. We show that the per‐release Gaussian mechanism—which clips and perturbs the aggregated pseudo‐gradient—is untrainable at any noise level that yields a meaningful privacy budget (the dispatch gap already exceeds 200% at ). Replacing it with per‐sample DP‐SGD, which clips each per‐sample gradient and perturbs their sum under subsampled‐Gaussian Rényi accounting, recovers training: at noise multiplier the composed privacy budget is () and the dispatch gap is , within the seed‐to‐seed standard deviation ( pp) of the no‐noise baseline .