Operating Envelopes of Synchronous and Asynchronous Federated Learning Under Communication Impairments in Smart-Grid Sensor Networks

Federated learning (FL) for distributed smart-grid monitoring is shaped jointly by statistical heterogeneity and communication timing. This study evaluates protocol operating envelopes rather than proposing a new aggregator. A frozen confirmatory campaign compares deadline-constrained FedAvg, FedAsync, FedBuff, and FedStaleWeight under ideal, intermittent, and degraded communication using equal client-dispatch and equal simulated-time budgets. In response to peer review, we add a separate physics-backed validation campaign using pandapower AC power flow, five independent physical trajectories, topology-aware IEEE 118-bus grouping, deadline sensitivity (0.5–5 s), an adaptive synchronous baseline, Deadline-FedProx, matched server-update controls, controlled Non-IID heterogeneity, a nonlinear multilayer perceptron, and buffer-, time-, horizon-, and tolerance-sensitivity analyses. The revised evidence shows that the numerical synchronous failure threshold is deadline-dependent rather than universal: under degraded IEEE 118 communication, fixed Deadline-FedAvg improves from 101.02 MW MAE at 1 s to 20.53 MW at 2 s and 17.57 MW at 5 s, while longer waiting increases schedule time. Adaptive deadlines recover many otherwise discarded updates, yet buffered asynchronous aggregation remains competitive or superior under severe communication and under a common cap on model-changing server updates. The qualitative mechanism persists across physical trajectories, Non-IID client distributions, and a nonlinear predictor. The results therefore support treating on-time update throughput, deadline policy, and aggregation semantics as first-class design variables, while the absolute operating boundaries remain specific to the evaluated communication and workload assumptions.