Optimizing Prosumer Policies in Periodic Double Auctions Inspired by Equilibrium Analysis

Bharat Manvi, Sanjay Chandlekar, Easwar SubramanianIJCAI 2024

Introduction

Periodic Double Auctions (PDAs) serve as the backbone for numerous modern markets, ranging from high-frequency financial trading to wholesale electricity grids. In these environments, participants often act as "prosumers"—entities that both buy and sell commodities to meet specific net procurement goals. For instance, an energy broker must fulfill customer demand by a set deadline, potentially selling excess renewable energy while purchasing the remainder from the market.

Optimizing bidding strategies in PDAs is notoriously difficult due to the sequential nature of the auctions and the incomplete information regarding competitors' needs. This paper addresses these challenges by deriving equilibrium policies and translating them into a practical, high-performance algorithm.

The Prosumer Challenge

Existing research often simplifies auctions into single-shot events or ignores the dual role of the prosumer. In a multi-round PDA, a bid placed in the first hour directly impacts the requirements for the remaining twenty-three. Agents must balance the immediate risk of high prices against the long-term uncertainty of future supply. This paper explicitly models the ability to place both "bids" (to buy) and "asks" (to sell), allowing brokers to strategically offload excess energy across rounds to minimize total costs.

Methodology: From Equilibrium to Algorithm

The authors approach the problem through a finite-horizon Markov game framework. The "state" of the game is defined by the wholesale supply curve and the prosumer's remaining requirements.

  1. Markov Perfect Nash Equilibrium (MPNE): By first assuming a "complete information" environment, the authors derive the MPNE for scenarios of both adequate and inadequate supply. In scarcity (inadequate supply), the equilibrium price naturally gravitates toward the "balancing price"—the penalty for failing to meet demand.
  2. MPNE-BBS Algorithm: To bridge the gap between theory and practice, the authors introduce Bidding Based on Supply (MPNE-BBS). Since real-world brokers cannot see competitors' private demands, MPNE-BBS uses historical orderbook data to reconstruct an approximate composite supply curve. It then applies the derived MPNE logic to estimate optimal limit prices and distributes bids to mitigate market volatility.

Experimental Evaluation

The researchers utilized the PowerTAC (Power Trading Agent Competition) simulator, a high-fidelity smart grid environment. The algorithm was tested across four demand levels—low, mid, high, and extreme—and compared against several state-of-the-art baselines:

  • VV21: A top-performing heuristic strategy.
  • DDPG: A Deep Reinforcement Learning approach.
  • SPOT: A strategy based on Monte Carlo Tree Search.

Key Results and Analysis

MPNE-BBS demonstrated remarkable consistency and efficiency across all tested scenarios. In head-to-head comparisons, it frequently outperformed both traditional heuristics and modern reinforcement learning models.

Performance comparison of MPNE-BBS against baselines across Low, Mid, High, and Extreme demand scenarios.

As shown in the evaluation metrics, MPNE-BBS consistently ranked as the best or second-best strategy. In "extreme demand" scenarios, the algorithm remained within 6% of the absolute theoretical best cost. In mid-to-high demand scenarios, it maintained a tight margin of 3.4% to 4.7% from the top-performing spot.

Evaluation of bidding strategies in high-volatility environments, highlighting the robustness of MPNE-BBS compared to RL-based DDPG.

A significant finding was the algorithm's superiority over Reinforcement Learning (DDPG) in extreme conditions. While RL agents often struggled with the high variance of scarcity, the equilibrium-inspired logic of MPNE-BBS allowed it to adapt more robustly to price spikes.

Conclusion

By grounding a practical bidding algorithm in formal equilibrium analysis, this research provides a robust framework for prosumers in complex markets. While the current model assumes relatively constant supply curves, future work aims to incorporate the volatility of renewable generation and move toward a Partially Observable Stochastic Game (POSG) model to better handle the inherent "fog of war" in competitive bidding.

Summary updated

BibTeX
@inproceedings{manvi2024optimizing,
  author = {Bharat Manvi and Sanjay Chandlekar and Easwar Subramanian},
  title = {{Optimizing Prosumer Policies in Periodic Double Auctions Inspired by Equilibrium Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure