Inverse optimization and robust aggregation based bidding strategy for distributed energy resource aggregators using multi-agent reinforcement learning

• A novel inverse optimization model surrogates the economic parameters of heterogeneous loads. • A two-stage adaptive robust optimization model to evaluate the flexibility of aggregator. • Hybrid framework embeds both physical models in a multi-agent bidding method for aggregators. Distributed energy resource aggregators are increasingly influential in electricity markets, yet bidding approaches based solely on machine learning or robust optimization often struggle with limited data, privacy concerns, and operation feasibility. We address these challenges with two physical-driven models that quantify an aggregator’s economic and technical characteristics through economic parameter surrogation and flexibility aggregation. First, a parametric inverse optimization model jointly learns the cost and constraint coefficients of responsive loads from historical observations, yielding reliable economic surrogates without restrictive assumptions. Second, a flexibility aggregation model constructs a feasible operation region that ensures both aggregation optimality and disaggregation feasibility. These models are embedded in a multi-agent bidding method so that learned policies respect identified costs, feasible flexibility, and robustness to uncertainty while relying only on aggregate information for privacy protection. Tests on the IEEE 118-bus system show that the proposed hybrid approach achieves a better fit to observed responses, maintains feasibility under uncertainty, and secures consistently higher revenues, while reducing dependence on large training datasets. The results highlight a practical path to reliable, privacy-aware bidding for aggregators in competitive markets.