Agent-Based Evolutionary Optimization for Resilient Power Flow Management in Distributed Smart Grids with High Renewable Penetration

Abstract

 Modern distribution grids are increasingly shaped by the variability of solar and wind generation, the spatial diversity of flexible loads, and the proliferation of fast power electronic interfaces that operate on time scales much shorter than legacy coordination processes. In this setting, centralized optimization alone struggles to provide timely decisions under uncertainty, while purely local control cannot guarantee global feasibility or systemwide resilience. This paper studies an agent-based evolutionary optimization framework for power flow management in distributed smart grids characterized by high renewable penetration and frequent fluctuations in net demand. The approach assembles autonomous agents at prosumer nodes, feeders, and substations that iteratively propose setpoints, exchange compact price and congestion signals, and adapt their strategies through continuous-time learning rules that respond to realized payoffs and observed reliability metrics. The design integrates physics-aware feasibility mappings that respect inverter, line, and voltage constraints, and it augments operational objectives with resilience terms that quantify performance degradation and recovery following disturbances. Market-inspired coordination with privacy-preserving message exchange aligns individual incentives with network-level feasibility and stability. Analytical results characterize convergence to feasible operating points under bounded uncertainty, and robustness to modeling mismatch and delayed communication. Numerical studies on radial and weakly meshed feeders with 60\%–90\% renewable share illustrate improvements in violation rates, recovery times, and loss of load probabilities when compared against baseline droop and static optimal power flow heuristics. The discussion highlights conditions under which the proposed dynamics remain well behaved, and identifies limitations arising from unmodeled device nonlinearities and adversarial behaviors.