⚡ electrical engineering

Customising Electricity Contracts at Scale with Large Language Models

This paper presents a scalable chat-based system that leverages Large Language Models integrated with functional power system analysis tools to enable end-users to negotiate and secure technically feasible, customized electricity contracts, thereby addressing labor shortages and improving grid capacity utilization.

Original authors: Jochen L. Cremer

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jochen L. Cremer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electricity grid as a massive, bustling highway system. Right now, getting a new car (a solar panel, an electric vehicle, or a factory) onto this highway is a nightmare of paperwork. You have to fill out rigid forms, wait months for an engineer to manually check if the road can handle your car, and often, you're told "no" or given a generic "one-size-fits-all" lane that doesn't quite fit your needs.

This paper proposes a revolutionary new way to handle this: A "Smart Traffic Cop" powered by AI that negotiates your entry onto the grid in real-time.

Here is the breakdown of the paper using simple analogies:

1. The Problem: The "Cookie-Cutter" Approach

Currently, the electricity system treats everyone the same.

The Analogy: Imagine a restaurant that only serves one fixed-size meal to every customer, regardless of whether they are a hungry teenager or a small child. If you want a custom meal, you have to wait for the chef (an engineer) to spend hours in the kitchen figuring it out. Because chefs are busy and short-staffed, they just say, "Here's the standard meal, take it or leave it."
The Reality: This leads to wasted capacity (the grid has room for more cars) or rejected connections (people can't plug in their EVs) because the system is too rigid to handle individual differences.

2. The Solution: The "AI Negotiator"

The authors built a chatbot system that acts like a super-smart, instant negotiator between you and the power grid.

The Analogy: Instead of filling out a form, you just text the "Traffic Cop." You say, "I want to charge my electric car at 5 PM."
The Magic: The AI doesn't just guess. It is connected to a digital twin of the actual power grid (a perfect video game simulation of the real wires and transformers).
- If your request is safe, the AI says, "Great! You're approved."
- If your request would cause a traffic jam (overload the grid), the AI doesn't just say "No." It says, "5 PM is too busy, but if you charge at 3 PM, it's a green light. Want to switch?"

3. How It Works: The "Translator"

The AI (Large Language Model) is great at talking human language, but bad at doing math. The power grid is great at math, but terrible at understanding human chat.

The Analogy: Think of the AI as a Translator and the Grid Software as a Calculator.
- You speak to the Translator: "I need power for my factory."
- The Translator converts your words into a specific command for the Calculator: "Run a safety check for Bus 5 with 2MW load."
- The Calculator does the heavy lifting and sends the result back to the Translator.
- The Translator tells you: "Your factory can run, but only if you turn off the ovens between 6 PM and 8 PM."
Why this matters: This ensures the AI never makes a dangerous mistake. It can chat, but it cannot break the laws of physics because the Calculator always double-checks the math before a contract is signed.

4. Three Real-World Scenarios Tested

The paper tested this "AI Negotiator" in three different situations:

Scenario A: The Homeowner (Low Voltage)
- The Situation: You want to charge your EV.
- The Result: By letting neighbors charge their cars at slightly different times (customizing the contract), the system could fit 7 times more electric cars into the same neighborhood without upgrading the wires. It turned a "traffic jam" into a smooth flow.
Scenario B: The Small Business (Medium Voltage)
- The Situation: A new shop wants to open and needs a lot of power. Usually, this takes months of meetings with engineers.
- The Result: The business owner chatted with the AI, which instantly told them, "Your location is too crowded for 2 PM, but perfect for 10 AM." They adjusted their opening hours, and the contract was signed in minutes, not months.
Scenario C: The Power Plant (High Voltage)
- The Situation: A power plant needs to shut down for maintenance. Usually, this is a complex scheduling puzzle that takes months to solve.
- The Result: The plant owner asked the AI, "Can we shut down next week?" The AI checked the grid and said, "No, that causes a blackout risk. But if you wait until week 8, it's safe." The scheduling was done instantly.

5. Safety and Security

The authors were very careful. They asked: "What if the AI lies or gets hacked?"

The Analogy: The AI is like a Receptionist, but the Security Guard (the functional programs) holds the keys. The Receptionist can talk to you and suggest things, but the Security Guard is the only one who can actually open the door. If the AI tries to suggest something unsafe, the Security Guard slams the door shut and tells the AI to try again. This ensures that even if the AI makes a mistake, the grid stays safe.

The Big Takeaway

This paper shows that we don't need to build more power lines to solve our energy problems. Instead, we can use AI to be more flexible. By treating every customer as unique and negotiating their energy use in real-time, we can unlock hidden capacity in the grid, save money, and speed up the transition to renewable energy.

It turns the electricity grid from a rigid, bureaucratic machine into a flexible, conversational partner.

1. Problem Statement

The modern electricity system is becoming increasingly complex due to the integration of massive numbers of heterogeneous end-users (households, SMEs, distributed generators) and the energy transition. The current contracting process faces two primary bottlenecks:

Grid Inefficiency due to Homogeneity Assumptions: To manage scale, Distribution System Operators (DSOs) and Transmission System Operators (TSOs) rely on standardized "block" contracts. These assume uniform consumption/injection patterns, leading to systematic overestimation of capacity needs and significant underutilization of grid infrastructure. They fail to leverage the hidden flexibility in diverse user behaviors.
High Transaction Costs and Delays: Customizing contracts to match specific user needs (e.g., time-of-use, specific outage windows) requires extensive manual engineering studies, legal review, and expert interaction. This process is slow (taking weeks to months), labor-intensive, and prone to bottlenecks, especially during engineer shortages. Consequently, many requests are rejected, or users are forced into suboptimal standard contracts, leading to overpaying or underutilizing capacity.

2. Methodology

The paper proposes a hybrid architecture that integrates Large Language Models (LLMs) with deterministic functional programs (APIs) to automate the design of technically feasible, customized electricity contracts.

Core Architecture

LLM as the Interface: The LLM acts as a conversational agent (chatbot) that interacts with non-expert users. It handles natural language understanding, intent recognition, and dialogue management.
Functional Programs as the Engine: The LLM does not perform power system calculations directly. Instead, it acts as a "function caller." It extracts parameters from user input and invokes pre-defined, deterministic functional programs ( $h(\alpha) \to \beta$ ) that interface with established power system simulation tools (e.g., AC power flow, SCOPF).
Safety & Validation Loop:
1. Extraction: The LLM extracts parameters (e.g., location, load profile, outage duration) from the chat context.
2. Execution: The functional program runs the physics-based grid analysis.
3. Validation: The program returns a binary feasibility result ( $\beta$ ) and specific constraints (e.g., "infeasible at hour 18").
4. Feedback: The LLM interprets the result and guides the user to adjust their request (e.g., shifting charging times) until a feasible contract is found.

Key Technical Components

Functional Programming: The system uses descriptive variable names (e.g., electricity_price instead of a123) to ensure the LLM can correctly map natural language to code arguments.
Deterministic Constraints: Physical feasibility (voltage limits, line loading, N-1 security) is enforced strictly by the deterministic solvers, not the LLM. This prevents the LLM from hallucinating unsafe grid conditions.
Error Handling: If the LLM extracts incorrect parameters, the deterministic function returns an error or warning, triggering a corrective feedback loop in the chat.

3. Key Contributions

The paper makes four primary contributions:

LLM-Chat with Functional Programs: Demonstrates high accuracy in using LLMs to access and orchestrate functional programs for power system analysis, effectively bridging natural language and engineering tools.
Scalable Customization for Households: Shows that the system is robust to user input variations (paraphrasing, typos) and can negotiate individualized time-of-use contracts for residential users at scale.
SME Self-Service: Enables Small and Medium Enterprises (SMEs) to self-study their connection requests via chat, with the system automatically verifying physical feasibility and grid security.
Chat-Enabled Maintenance Planning: Demonstrates the application for TSOs to manage transmission equipment outages, offering real-time alternative scheduling to asset owners while maintaining N-1 security.

4. Results and Case Studies

The study was validated using three distinct use cases across different voltage levels:

Case 1: Residential Low-Voltage (LV) Grids
- Scenario: Negotiating EV charging contracts for 15 households with heterogeneous preferences.
- Result: By customizing contracts to exploit heterogeneity (shifting charging times), the system allowed 7 times more EVs to connect to the same grid capacity compared to standard block contracts.
- Robustness: The system maintained high accuracy even with user input variations (paraphrasing) and typographical errors (up to ~2% error rate), though performance degraded significantly with higher error rates.
Case 2: Industrial Medium-Voltage (MV) Grids (Oberrhein 179-bus system)
- Scenario: An SME developer requesting a 2MW connection.
- Result: The chatbot iteratively guided the user to adjust their load profile (e.g., shifting commercial hours) to avoid voltage violations at specific buses.
- Efficiency: What typically takes weeks of manual engineering review was resolved in minutes with minimal human oversight.
Case 3: Transmission High-Voltage (HV) Grids (IEEE 118-bus system)
- Scenario: Planning generator maintenance outages (N-1 security constrained).
- Result: The system solved 365 Security-Constrained Optimal Power Flows (SCOPFs) in ~1 minute. It successfully identified that a requested outage (Week 3) was infeasible due to simultaneous maintenance on a nearby generator and proposed alternative feasible windows (e.g., Week 8).
- Privacy: The study analyzed security risks, concluding that while brute-force attacks could map grid congestion patterns, the system structure (hiding internal data) limits the exposure of sensitive grid data to the user.

5. Significance and Impact

Economic Efficiency: Drastically reduces transaction costs for contract customization, making it economically viable to offer flexible, individualized contracts to millions of users.
Grid Optimization: Unlocks hidden flexibility in the grid by moving away from rigid standard contracts, allowing for higher penetration of renewable energy and EVs without requiring immediate infrastructure upgrades.
Operational Speed: Reduces planning horizons from months to minutes for connection requests and outage management, addressing the bottleneck of engineer shortages.
Safety Assurance: By decoupling the LLM (probabilistic) from the physics engine (deterministic), the system ensures that no physically infeasible or unsafe contracts are issued, addressing a critical safety concern in AI applications for critical infrastructure.

Conclusion

This paper establishes a foundational framework for using LLMs not just as chatbots, but as orchestrators of engineering workflows. It proves that natural language interfaces can effectively manage complex power system constraints, enabling a shift from standardized, inefficient grid interactions to scalable, customized, and flexible energy contracts.