Multi-agent coordination and uncertainty adaptation in deep learning–assisted hierarchical optimization for renewable-dominated distribution networks

The rapid proliferation of renewable energy sources has introduced multi-layered uncertainty into modern power system operation, challenging conventional deterministic and stochastic optimization frameworks. To address this complexity, this study proposes a deep learning-assisted distributionally robust optimization (Deep-DRO) framework designed to enhance both economic efficiency and operational reliability under uncertainty. The model integrates a hierarchical coordination architecture, wherein deep learning modules infer the probabilistic structure of uncertain variables-such as solar irradiance, wind availability, and load fluctuation-while the DRO layer enforces system-wide robustness through an adaptively reshaped ambiguity set. The learning-assisted ambiguity reconstruction enables the optimization to dynamically adjust conservativeness, improving the tradeoff between cost, reliability, and renewable utilization. Methodologically, the proposed framework employs a multi-agent dispatch structure consisting of three decision layers-county, feeder, and distributed energy resource (DER)-each learning distinct policy mappings through reinforcement-guided coordination. Deep networks trained on high-resolution meteorological and operational data estimate scenario distributions, while the robust optimization core minimizes expected cost and reliability penalties under distributional ambiguity. The resulting hybrid system seamlessly couples data-driven forecasting and model-based optimization, bridging the gap between predictive intelligence and operational robustness. To ensure scalability and interpretability, convergence diagnostics, sensitivity analyses, and cost decomposition studies are performed across multiple test systems and uncertainty scenarios. Simulation results on a benchmark multi-region distribution network demonstrate substantial performance gains. Compared to conventional DRO, the Deep-DRO model reduces total operational cost by 11.0-13.5%, improves reliability indices from 0.864 to 0.911, and raises renewable utilization from 85.6% to 89.7%. The integrated deep learning mechanism effectively captures latent correlations among stochastic parameters, enabling the system to maintain resilience even under 30% higher uncertainty variance. Furthermore, carbon emissions decline by 28.6% relative to baseline, confirming that the proposed method achieves an intrinsic balance between economic optimization and environmental sustainability. The analysis reveals that hierarchical learning fosters adaptive coordination among agents, while the robust layer guarantees performance consistency across uncertain conditions. The study thus advances a generalizable paradigm for intelligent, risk-aware energy management, offering theoretical and practical implications for future power system restoration, smart grid autonomy, and sustainable dispatch design.