Behavioral Generative Agents for Power Dispatch and Auction

This paper demonstrates that large language model-powered generative agents, enhanced by in-context learning, can effectively relax the rigidity of traditional mathematical models to replicate both rational strategies and systematic behavioral deviations in power dispatch and auction scenarios.

The Gist

Imagine you are trying to teach a very smart, but sometimes overly literal, robot how to make decisions in the messy, unpredictable real world. Specifically, you want to know: Can a robot powered by a Large Language Model (like the one writing this response) act like a real human when managing electricity or bidding in an auction?

This paper says, "Yes, but with a little help."

Here is the story of their experiment, broken down into simple concepts and analogies.

The Big Idea: "Homo Silicus"

Usually, when engineers model how people use electricity or buy things, they assume everyone is a perfectly rational robot. They assume everyone always tries to make the most money possible and never makes a mistake.

But real humans aren't like that. Sometimes we panic during a blackout. Sometimes we get greedy in an auction. Sometimes we care more about safety than profit.

The authors wanted to build a "Homo Silicus" (a silicon-based human). They used AI agents to simulate these messy human behaviors to see if the AI could learn to act like us, not just like a calculator.

---

Experiment 1: The Home Battery & The Blackout

The Setup:
Imagine you have a home battery. You can charge it when electricity is cheap and sell it back when it's expensive to make a profit. This is like a video game where the goal is usually to get the highest score (money).

The Twist:
Suddenly, the power grid goes down (a blackout). Now, your battery isn't just a money-maker; it's your lifeline. You need to keep enough charge to keep your lights on, even if it means you make less money.

The Problem:
Standard math models (the "perfect robots") would tell you to drain the battery completely to sell the last drop of energy for profit, even if it leaves you in the dark during a blackout. They don't "feel" fear.

The Solution (The "Study Buddy" Trick):
The researchers used a "smart" AI (let's call him Professor AI) to figure out the right behavior during a blackout. Professor AI realized: "Hey, I should save some battery for emergencies!"

Then, they used a technique called In-Context Learning (ICL). Think of this as giving a student (a smaller, cheaper AI) a cheat sheet or a storybook written by Professor AI.

• Without the cheat sheet: The student AI acts like a greedy robot, draining the battery.
• With the cheat sheet: The student AI reads the story, sees the example of saving power for the blackout, and suddenly acts like a cautious human. It keeps a "safety reserve."

The Result:
By showing the AI examples of "good behavior" (like saving energy for an emergency), they could teach a smaller AI to prioritize safety over profit, just like a real human would.

---

Experiment 2: The Power Auction

The Setup:
Imagine a high-stakes auction for "Grid Access Rights" (permission to plug your data center into the power grid). There are two companies bidding. The auction goes up in rounds, like a silent auction where the price goes up every time someone bids.

The Players:
The researchers created three different types of AI bidders to see how they behaved:

1. The Rule-Follower: This AI just follows the rulebook. "If I want it, I bid." It doesn't think ahead.

   • Result: It got too aggressive. It kept bidding even when it was losing money, just because it wanted to "win" the item. It was like a kid at a toy store screaming, "I WANT IT!" without checking their wallet.

2. The Short-Sighted Greedy Guy: This AI only cares about making money right now.

   • Result: It acted exactly like a perfect math model. It bid the minimum necessary to win, calculated the profit, and stopped. Very boring, very rational.

3. The Strategic Mastermind: This AI plays the long game. It thinks, "If I bid a little higher now, I might scare off my opponent and win the whole auction later."

   • Result: This AI was the most interesting. It bid aggressively early on to secure a dominant position, but it stopped before it went broke. It balanced aggression with profit, just like a savvy human business owner.

The Magic Ingredient:
The key to making the "Mastermind" work was prompt engineering. The researchers gave the AI a specific "personality" and a "journal" to write in after every round.

• The Journal: After every bid, the AI had to write down: "What did I do? Why did I do it? What will I do next?"
• This forced the AI to think before it acted, preventing it from making random, crazy bids. It made the AI feel like it had "experience."

---

Why Does This Matter?

Think of the power grid as a giant, complex orchestra.

• Old way: We try to conduct the orchestra using a rigid sheet of music (math models) that assumes every musician plays perfectly.
• New way: We use AI agents to simulate what happens if the musicians are tired, scared, or trying to show off.

The Takeaway:
This paper proves that we can use AI to create "Digital Humans" for energy systems.

1. We can teach them: By showing them examples (In-Context Learning), we can make them act more responsibly during emergencies (like blackouts).
2. We can test policies: Before we change real electricity laws, we can run simulations with these AI humans to see if the new rules will cause chaos or work smoothly.

In short, we are moving from building calculators that do math, to building simulations that understand human behavior. And that is a huge step toward a smarter, safer energy future.

Technical Summary

Here is a detailed technical summary of the paper "Behavioral Generative Agents for Power Dispatch and Auction."

1. Problem Statement

Traditional energy system models rely on mathematical optimization frameworks (e.g., Dynamic Programming) that assume agents are fully rational, maximizing well-defined objectives like cost minimization or utility. However, real-world human decision-makers exhibit bounded rationality, heterogeneous preferences, and context-dependent behaviors (e.g., prioritizing reliability over profit during blackouts, or strategic aggression in auctions).

Current Large Language Model (LLM) agents show promise as "homo silicus" (simulated humans) but face challenges in real-world applications, including hallucinations, output variability, and difficulty in maintaining consistent strategic reasoning. The paper addresses the need for a flexible testbed to model these complex human behaviors in power dispatch and auction mechanisms without the high cost and logistical constraints of human-subject experiments.

2. Methodology

The authors propose a Behavioral Generative Agent framework powered by LLMs, utilizing In-Context Learning (ICL) and a structured reasoning architecture called TARJ (Thought-Action-Reflection-Journal). The study consists of two proof-of-concept experiments:

A. Experimental Setup 1: Home Battery Management (Power Dispatch)

• Scenario: A 20-day simulation of a home battery with stochastic electricity prices ($5 or $10/kWh) and a one-day blackout intervention.
• Agent Architecture:
  • TARJ Framework: The LLM is prompted to generate a Thought (reasoning), Action (charge/discharge/hold), Reflection (outcome analysis), and Journal (memory storage). The Journal is fed back into subsequent prompts to enable adaptive learning.
  • ICL Module: To transfer behavioral patterns from a stronger model to a smaller, cheaper one, the authors use o1-preview (strong reasoning model) to generate example decision trajectories during blackouts. These examples are injected as prompts for gpt-5-mini (smaller model).
• Benchmarks: Decisions are compared against an optimal Dynamic Programming (DP) policy (rational arbitrage) and a Greedy heuristic.
• Personas: Three distinct agent personas are tested: Thinker, Realist, and Feeler (prioritizing reliability).

B. Experimental Setup 2: Power Network Access Auctions (SAA)

• Scenario: A Simultaneous Ascending Auction (SAA) where two bidders compete for grid withdrawal rights at two locations (A and B).
• Agent Architectures: Three distinct LLM agents are designed with different strategic horizons:
  1. Rule-Centric: Bids strictly based on auction rules and private valuations without strategic foresight.
  2. Myopic-Profit: Optimizes for immediate utility in the next round only.
  3. Strategic-Outcome: Optimizes for terminal utility (long-term profit), potentially sacrificing short-term gains to secure a dominant market position.
• Mechanism: Agents use the TARJ framework with ICL to maintain consistency across rounds. The auction ends when no prices or high bidders change.
• Baseline: Compared against the Straightforward Bidding Strategy (a standard economic benchmark where bidders maximize immediate surplus).

3. Key Contributions

1. Behavioral Transfer via ICL: Demonstrated that behavioral patterns discovered by a high-capacity reasoning model (o1-preview) can be effectively transferred to smaller, cost-efficient models (gpt-5-mini) using example-based prompting. This allows smaller agents to exhibit sophisticated behaviors (e.g., prioritizing energy reserves) without retraining.
2. Structured Reasoning for Strategic Consistency: The TARJ framework (specifically the Journal component) successfully stabilizes agent behavior across multiple rounds, preventing erratic bidding and enabling coherent long-term planning in auctions.
3. Modeling Heterogeneous Preferences: The study successfully simulated diverse human-like behaviors, from profit-maximizing rationality to risk-averse reliability seeking and strategic aggression, which are difficult to capture in rigid mathematical models.
4. Proof-of-Concept Testbed: Established a scalable, low-cost methodology for evaluating energy policies and market designs using "homo silicus" before deploying them in real-world systems.

4. Key Results

Battery Management Results

• Blackout Response: Without ICL, agents generally followed profit-maximizing logic, leaving the battery with low State of Charge (SoC) at the end. With ICL examples, agents (specifically Feeler and Realist personas) learned to fully discharge during the blackout and retain higher energy reserves afterward, deviating from the optimal DP policy to prioritize reliability.
• Persona Differentiation: ICL enhanced the ability of smaller models to distinguish between personas. Agents with ICL showed increased behavioral heterogeneity, clearly reflecting their specific preferences (e.g., Feeler retaining more energy than Thinker).
• Limitation: While qualitative patterns transferred well, quantitative differences remained (e.g., slight variations in terminal rewards), indicating bounds to ICL-based transfer.

Auction Results

• Myopic-Profit vs. Straightforward: The Myopic-Profit agent's bidding trajectory closely aligned with the theoretical Straightforward Bidding Strategy, confirming that LLMs can replicate rational, short-term economic optimization.
• Strategic Aggression: The Strategic-Outcome agent exhibited more aggressive bidding in early rounds compared to the Myopic-Profit agent and the baseline. It utilized higher bids to secure a dominant position for desired products, demonstrating long-term strategic planning.
• Rule-Centric Irrationality: The Rule-Centric agent (lacking strategic CoT guidance) displayed irrational escalation, continuing to bid even after competitors withdrew. This highlights the necessity of structured reasoning (CoT/ICL) to prevent hallucinated or suboptimal behaviors.
• ICL Effectiveness: The ICL modules successfully prevented erratic bidding in the Strategic-Outcome agent, ensuring it remained focused on terminal profitability rather than "winning at any cost."

5. Significance and Future Work

• Policy & Design: This work provides a tool for energy policymakers to simulate population-level human responses to new tariffs, blackouts, or market designs, offering insights that pure optimization models miss.
• Human-AI Collaboration: It suggests a future where LLM agents act as intermediaries between rigid mathematical models and costly human trials, allowing for rapid iteration of energy system designs.
• Future Directions: The authors note that current experiments are stylized. Future work aims to integrate real-world market datasets, complex grid reliability constraints, and eligibility rules (currently omitted to reduce hallucination) to further validate the agents' robustness.

In conclusion, the paper validates that LLM-powered generative agents, when augmented with In-Context Learning and structured reasoning frameworks, can effectively model complex, heterogeneous human decision-making in critical energy systems, bridging the gap between theoretical optimization and practical human behavior.