⚡ electrical engineering

Physically Constrained Agentic AI for Energy Scheduling

This paper presents a hierarchical ReAct Energy Management System that separates physical authorization from language generation to enable agentic AI to generate feasible, cost-effective day-ahead energy schedules for appliances, EVs, and thermal control, achieving high success rates and near-optimal costs while strictly adhering to physical constraints.

Original authors: Dafnag Zhao, Yang Deng, Zhengmao Li

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dafnag Zhao, Yang Deng, Zhengmao Li

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 a world where your home doesn't just obey commands but understands your desires. You might tell your house, "I want the laundry done, the car charged, and the living room cozy, but please don't make the electricity bill skyrocket." This is the dream of Agentic AI: a smart system that doesn't just chat with you but actually does things, breaking big goals down into small, coordinated actions. However, there's a catch. While a chatbot is great at writing a poem, it's terrible at physics. If an AI guesses that your heater can run at full blast while your washing machine is spinning and your electric car is charging, it might accidentally blow a fuse or melt a wire. This is the challenge of Energy Management: balancing what you want with the hard, unyielding laws of electricity and heat. Scientists are trying to build systems that can talk to you in plain English but still respect the strict math of the power grid, ensuring that your home stays safe and efficient without needing a human engineer to double-check every decision.

This paper introduces a clever new way to solve that problem, acting like a "smart butler" with a very strict "safety inspector" standing right next to them. The authors built a system called a Hierarchical ReAct Energy Management System. Think of it as a team of specialists: one agent handles the washing machine, another the electric car, and a third the heating system. A central "orchestrator" listens to your request and asks these specialists to come up with a plan. But here is the magic trick: the AI is allowed to dream up ideas and revise them, but it is never allowed to turn the switches on its own. Before any plan is executed, a "deterministic critic"—a rigid, math-only computer program—steps in. This critic doesn't care about polite conversation; it only checks the numbers. It reconstructs the entire day's schedule, simulates the heat flow, checks the power limits, and ensures no appliance is turned on when it shouldn't be. If the plan passes this strict test, it gets authorized; if it fails, it gets rejected, and the AI has to try again.

The researchers tested this system using a powerful AI model (specifically, checkpoints of the Qwen 3.5 series) against a standard mathematical solver (Pyomo/Gurobi) to see if the AI could actually handle the complex math of scheduling. They set up a scenario where the home needed to manage a washing machine, a dishwasher, an electric vehicle, and a heat pump, all while keeping the indoor temperature between 22 and 24°C during the day and staying under a 12-kW power limit.

The results were a mix of "not bad" and "pretty good," depending on the size of the AI brain. When the AI was asked to schedule just one appliance, it succeeded about 83.3% of the time. However, when the task got harder and required coordinating everything at once (the "coupled" task), the smaller AI models got stuck or timed out. But the larger models, specifically the 27B and 35B-A3B versions, managed to pass the safety inspector's test in all 6 out of 6 attempts after a few rounds of feedback. The 35B-A3B model created a schedule that cost 1592.697 JPY, which is slightly higher than the perfect mathematical optimum of 1343.380 JPY, but it was a feasible plan that respected all the physical rules. The 27B model also succeeded but was more expensive at 2522.499 JPY.

Crucially, the paper emphasizes that this success is conditional. The system works because the "critic" acts as a fail-safe. If the AI suggests something impossible, the critic catches it. The authors found that without this strict separation between the "creative" AI and the "mathematical" checker, the system would likely produce plans that look good in text but fail in reality. They also noted that this was a simulation using specific weather and price data from August 1, 2026, and the system doesn't yet account for real-world messiness like humidity, battery degradation, or sudden power outages.

In short, this paper shows that we can build AI agents that talk to us in natural language and manage our homes, but only if we keep a strict, unblinking mathematical watchdog on the leash. The AI can propose, but the math must dispose. The larger AI models showed promise in navigating this tightrope, finding schedules that were safe and functional, even if they weren't quite as cheap as the perfect mathematical solution. It's a step toward homes that understand us, but it proves that for energy, the rules of physics still have the final say.

Technical Summary: Physically Constrained Agentic AI for Energy Scheduling

Problem Statement
Day-ahead energy scheduling requires integrating household intent (expressed via natural language) with complex physical constraints, including electricity tariffs, weather, calendar limits, and device models. While Large Language Models (LLMs) can extract parameters from dialogue and agentic systems can delegate decisions to specialists, a critical gap remains: a linguistically plausible schedule does not guarantee satisfaction of coupled electrical and thermal constraints. The central challenge is authorization—ensuring that individually plausible outputs from specialist agents do not conflict when integrated (e.g., shared feeder limits or common thermal states). Current approaches often lack a mechanism to rigorously verify that a proposed schedule adheres to the physical reality of the system before execution.

Methodology
The paper proposes a Hierarchical ReAct Energy Management System (EMS) that separates language generation from physical authorization. The system operates through the following components:

Orchestrator and Specialists: A ReAct-based orchestrator translates natural-language requests into coordinated actions. It delegates subproblems to three specialist agent types:
- Shiftable Appliances: Handles binary start times for uninterrupted cycles (e.g., washing machines, dishwashers).
- EV Charging: Manages charging cycles.
- Thermal Control: Generates variable power profiles for heat pumps.
Deterministic Critic (Fail-Closed Policy): A critical innovation is the separation of proposal and authorization. Agents may propose and revise actions, but only a deterministic critic can release the final candidate. The critic performs:
- Schema Validation: Checks typed representations and reconstructs binary cycles.
- Physical Simulation: Propagates thermal states using a linear model ( $T_{t+1} = a_m(1000p_t) + b_m T_t + c_m T^{out}_t + d_m$ ) and calculates aggregate power ( $P^{tot}_t$ ).
- Constraint Checking: Verifies comfort bounds (22–24°C during occupied windows), device power limits, and whole-home feeder limits (12 kW).
- Sign-Safe Feedback: If constraints are violated, the critic calculates feasible intervals for power adjustments. If an interval is empty, the candidate is rejected; otherwise, it requests a complete resubmission of the trajectory.
Experimental Setup: The system was evaluated using Qwen 3.5 checkpoints (ranging from 0.8B to 35B-A3B parameters) against mathematical optimization references (Pyomo/Gurobi). The evaluation included:
- MIP Tasks: Single-appliance scheduling.
- LP Tasks: Heat-pump-only scheduling under various weather/price scenarios.
- MILP Tasks: Whole-home coupled scheduling (appliances + heat pump + feeder limit).
- Constraints: A standard "occupied-window" policy enforced comfort from 09:00–18:00 and required zero heat-pump power post-window.

Key Contributions

Hierarchical ReAct EMS: A framework mapping natural language to coordinated specialist actions with a configurable LLM backend.
Fail-Closed Authorization: A critic mechanism that enforces physical constraints through sign-safe thermal feedback, coupled multi-step revision, and explicit rejection. It ensures that only candidates passing a fresh, independent replay are authorized.
Empirical Evaluation: A rigorous comparison of discrete, continuous, and coupled scheduling tasks across six Qwen 3.5 checkpoints, demonstrating that agentic workflows can achieve feasibility under a declared physical model when paired with deterministic verification.

Results

Single-Appliance (MIP): Achieved 83.3% feasibility and 33.3% exact optimality across 54 runs.
Coupled Scheduling (MILP):
- Initial Attempts: No initial whole-home candidates satisfied all binary, thermal, and feeder checks simultaneously.
- Localized Feedback: Produced no accepted coupled schedule.
- Multi-Step Policy: A multi-step revision policy authorized 6/6 current runs (3 for 27B, 3 for 35B-A3B).
- Cost Efficiency:
  - 35B-A3B: Achieved a cost of 1592.697 JPY (18.56% above the mathematical optimum of 1343.380 JPY).
  - 27B: Achieved a cost of 2522.499 JPY (87.77% above the optimum), primarily due to suboptimal appliance timing and higher heat-pump usage.
- Constraint Adherence: All accepted schedules satisfied the 12-kW feeder limit, uninterrupted appliance cycles, and the requirement for zero heat-pump energy after 18:00.
Latency: The actor-critic pipeline averaged 33.92s (35B-A3B) and 84.84s (27B), compared to 0.0487s for the mathematical solver.

Significance and Claims
The paper establishes a fail-closed workflow for agentic Mixed-Integer Programming (MIP) and Mixed-Integer Linear Programming (MILP) energy scheduling. The primary significance is the demonstration that agentic AI can coordinate complex, coupled decisions while remaining strictly bound by a deterministic physical model.

The authors explicitly state that:

Authorization is Conditional: Acceptance is valid only under the encoded schema, thermal surrogate, device limits, and integrated feeder constraint. The system does not validate omitted dynamics (e.g., humidity, battery state of charge, equipment cycling) or infer missing constraints from LLM explanations.
Robustness Limitations: The experiment does not establish robustness to model error. Post-hoc perturbations reduced feasibility, indicating that operational deployment would require calibrated uncertainty sets, receding-horizon feedback, and equipment-specific constraints.
Role of Optimization: Mathematical programming serves as a numerical reference, not a replacement for the agentic layer. The agentic workflow provides intent interpretation and revision, but the deterministic critic remains the sole gatekeeper for execution.

The work concludes that while the current results are promising for controlled, day-ahead scheduling, broader operational claims require further validation against field measurements and uncertainty quantification.

Technical Summary: Physically Constrained Agentic AI for Energy Scheduling

More like this