PowerAtlas: Feasibility-Aware Language Models for Electricity-Computing Co-Scheduling

arXiv:2607.26710 · cs.LG · Submitted 2026-07-29 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "PowerAtlas: Feasibility-Aware Language Models for Electricity-Computing Co-Scheduling".

Jane: The gist The rapid growth of AI workloads is turning data centers into largescale, volatile, yet spatiotemporally flexible grid loads, creating an urgent need for coordinated electricity-computing scheduling.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, let's talk about what PowerAtlas actually does in terms of its structure and its authors. This paper is about making LLMs capable of handling complex physical constraints in electricity-computing co-scheduling.

Jane: The title itself points to the main challenge: combining power systems knowledge with task scheduling decisions into one coherent output > Page one #pg1.

Lu: The authors are tackling this by framing it as a single decision problem for an LLM agent, which is a significant shift from how these domains have been handled before > Page three #pg3.

Meng: It’s interesting to see the authors define the input specification S with all those components—the constraints, the task description, and domain knowledge—all feeding into one decision D > Page three #pg3.

Tom: They are essentially showing how to bridge that gap between natural language intent and those strict physical constraints of a coupled system > Page three #pg3.

Jane: This moves beyond just having the model generate text in the right format; it demands physical feasibility from the very first generation > Page two #pg2.

The paper's summary: Tom: Now, let's break down how PowerAtlas actually works internally. It focuses on making this joint decision process robust and effective across different models.

Jane: One major component they introduce is Feasibility-Aware Group Relative Policy Optimization, or FA-GRPO > Page two #pg2.

Lu: This stage aligns the policy by treating the decision generation as a token-level Markov decision process, optimizing a group relative objective J(theta) > Page two #pg2.

Meng: What does that mean practically for deployment? Does it make the model less brittle when faced with unexpected load changes in a data center environment?

Tom: The reward function they use is structured to encode "satisfy hard constraints first, then reward quality," where quality is measured by coupling the two domains into one scalar value > Page two #pg2.

Jane: That coupling of compute and grid quality into a single reward function is a very smart way to balance those two separate objectives during training > Page two #pg2.

Lu: Specifically, the grid quality term folds constraint violations into an equivalent cost where Qgrid equals min(one c star over (c theta plus C LOL Viol)) > Page two #pg2.

Tom: So they are literally turning constraint violations into a measurable cost that the policy has to minimize while also maximizing compute reward > Page two #pg2.

Meng: I see how that helps with the economic efficiency side, tying it directly into quantifiable penalties for grid issues > Page two #pg2.

The paper's improvements: Jane: Let's look at the specific technical improvements they suggest in PowerAtlas. They focus on how they make this approach better than previous attempts at solving electricity-computing co-scheduling.

Tom: A major improvement is the deterministic verification step, which ensures adherence to physical constraints by parsing the block and replaying it through an hourly nodal DC feasibility linear program > Page two #pg2.

Lu: This verifier is what makes sure that the schedule is physically feasible and economically efficient at the same time > Page nine #pg9.

Meng: That sounds like a very rigorous check. How does that compare to just having a basic sanity check?

Tom: It's much stricter than that. It ensures decisions are not just format-valid but actually meet the physical and economic criteria > Page two #pg2.

Jane: And they map the input specification S to the joint decision D by decoding it autoregressively as a single templated token sequence y > Page three #pg3.

Lu: So, instead of two separate models making two separate decisions, it’s one model generating one sequence for both parts of the schedule > Page two #pg2.

Conclusion: Tom: So, we've seen how PowerAtlas tackles electricity and computing scheduling as one big decision for an AI agent by combining supervised learning with feasibility-aware policy optimization.

Jane: It’s about giving the model more context than just a price signal; it needs to understand the physical boundaries of what's possible > Page three #pg3.

Lu: The paper treats electricity-computing co-scheduling as one single decision problem for an LLM agent, and they use warm-starting on solved cases to teach the model the output format and basic feasible patterns > Page two #pg2.

Meng: So it’s about giving the model more context than just a price signal; it needs to understand the physical boundaries of what's possible > Page three #pg3.

Tom: And they use Feasibility-Aware Group Relative Policy Optimization, or FA-GRPO, to align it toward that joint objective > Page four #pg4.

Jane: This stage aligns the policy by casting decision generation as a token-level Markov decision process, optimizing a group relative objective J(theta) > Page two #pg2.

Meng: I'm curious about the verification part. How does it check if the schedule is actually physically possible when you’re dealing with something as complex as power flow?

Tom: That’s where they use a deterministic verifier that parses the block and replays it through an hourly nodal DC feasibility linear program > Page two #pg2.

Lu: And this process involves mapping the input specification S to a joint decision D by decoding it autoregressively as a single templated token sequence y > Page three #pg3.

Jane: It’s about bridging that gap between natural language intent and those very strict physical and scheduling constraints of the coupled system > Page three #pg3.

Meng: When you think about deploying this, what are your main concerns? What breaks when you try to run this kind of complex reasoning at scale?

Tom: The paper shows that even across eleven different large language models, PowerAtlas works well under realistic physical operating conditions > Page one #pg1.

Lu: The results show that the violation falls by eighty-four to ninety-eight percent, and the cost drops by up to twenty-six times across those three different backbones > Page one #pg1.

Meng: Twenty-six times reduction in cost sounds substantial if it holds up on real traffic, but I still wonder about the data source—how realistic is that ECBench dataset?

Jane: To summarize, PowerAtlas shows we can tackle electricity-computing co-scheduling as one single problem for an LLM agent by combining supervised learning with feasibility-aware policy optimization > Page nine #pg9.

Lu: It’s a unified approach that moves away from treating electricity as just an exogenous price for compute > Page three #pg3.

Meng: For us at the startup level, this suggests that if we can get the warm start right, you can use FA-GRPO to quickly dial in a feasible schedule without needing exhaustive manual tuning for every new scenario > Page one #pg1.

Tom: And the online deployment experiment confirmed this; they built an experimental electricity-computing network across three data centers in one province of China where the agent read live facility state as the specification S > Page six #pg6.

Jane: That’s how it moves from theory to practice, showing that power-side schedules are checked against a real model rather than just being directly actuated > Page six #pg6.

Lu: This work is really about how to make these LLM agents not just format-valid, but actually physically feasible and economically efficient at the same time > Page nine #pg9.

Meng: It’s a step toward deploying AI workloads on real infrastructure with better coordination between power and compute > Page six #pg6.

Jane: That’s the gist of PowerAtlas: by combining those stages, we can tackle these hard physical constraints in one go > Page nine #pg9.

Tom: We've seen how PowerAtlas uses a deterministic verifier to ensure schedules are physically feasible and economically efficient at the same time > Page two #pg2.

Lu: It’s about making sure the agent produces decisions that aren't merely format-valid but physically feasible and economically efficient at the same time > Page nine #pg9.

Meng: That means we need to focus on building those deterministic verifiers if we want to deploy this reliably without constant manual overrides > Page six #pg6.

Jane: Exactly. And that’s where the next big challenge lies for integrating complex AI reasoning into physical systems > Page nine #pg9.

Kaiwen Jiang, Siya Xu, Ziyue Zhu, Chao Yang, Anh Tuan Luu, Haoran Luo

Beijing University of Posts and Telecommunications · State Grid Liaoning Electric Power Company, Ltd. · Nanyang Technological University

cs.LG

Submitted: 2026-07-29

Updated: 2026-10-15

Comments: 28 pages, 9 figures, 16 tables. Code: https://github.com/JAVA-Jiang/PowerAtlas

Code: https://github.com/JAVA-Jiang/PowerAtlas

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 85/100

The gist: The gist The rapid growth of AI workloads is turning data centers into largescale, volatile, yet spatiotemporally flexible grid loads, creating an urgent need for coordinated electricity-computing

Key concepts

ECCS Task: This is the main goal: generating a single decision that simultaneously satisfies two competing constraints—grid security (keeping power stable) and task service (meeting compute demands). The agent must do this while respecting physical grid models, compute needs, and domain knowledge.
ECBench Dataset: This dataset contains 2,000 complex scenarios where the optimal joint decisions have already been calculated using a Mixed-Integer Linear Program (MILP). It is used to train and evaluate the agent because it provides realistic data anchored by real data center specifics and task distributions.
PowerAtlas Framework: This is a model-agnostic system that guides an LLM to output a joint decision (power distribution and compute schedule) in one step. It uses warm-starting on solved instances to teach the LLM the correct output format before optimizing its behavior through reinforcement learning.
Feasibility-Aware Policy Alignment (FA-GRPO): This is the core optimization stage where the agent learns. It treats decision generation like a game, rewarding actions that prioritize satisfying hard grid constraints first, and then rewards overall quality by balancing compute performance and grid cost.

Terminology

Benchmark — 1× in this paper · explained in 2 episodes
Large Language Models (LLMs) — 1× in this paper · explained in 8 episodes
Physical Constraints — 1× in this paper · explained in 7 episodes

Summary

The gist The rapid growth of AI workloads is turning data centers into largescale, volatile, yet spatiotemporally flexible grid loads, creating an urgent need for coordinated electricity-computing scheduling.

ECCS Task

The ECCS task is defined as generating a joint decision that simultaneously satisfies grid-security and task-service constraints given the grid physical model, the compute-task demands, and domain knowledge. This task bridges operational intent expressed in natural language and the strict physical and scheduling constraints of the coupled system so that the agent produces decisions that are not merely format-valid but physically feasible and economically efficient at the same time, and does so in a single pass over the specification. The input specification S comprises four parts (Figure 3), including grid physical constraints SC, task description ST, retrieved domain knowledge SK, and output-format contract SF<ref:2607.26710#pg4>

ECBench Dataset

The ECBench dataset is constructed by a four-stage pipeline that fuses three data sources into 2,000 coupled scenario instances all solved to optimality. Realism is anchored on the data-center side through rack count calibration from a real data center, while task-level distributions are synthesized and deliberately relaxed in schedulability so that instances remain non-trivially solvable<ref:2607.26710#pg5> Each instance is solved to optimality by a mixed-integer linear program (MILP) with Gurobi, yielding the reference decision D and reference cost c used for both supervised initialization and reward normalization.

PowerAtlas Framework

The PowerAtlas framework combines supervised initialization on solved instances with physical verification and Feasibility-Aware Group Relative Policy Optimization (FA-GRPO) into a model-agnostic plug-in. The agent maps the input specification S to a joint decision D = (DP, DC) by decoding it autoregressively as a single templated token sequence y. This process involves warm-starting on compliant, optimally solved instances to teach the output format and basic feasible decision patterns

Feasibility-Aware Policy Alignment (FA-GRPO)

The FA-GRPO stage aligns the policy by casting decision generation as a token-level Markov decision process, optimizing a group-relative objective J(θ). The reward encodes “satisfy hard constraints first, then reward quality,” with the quality term coupling the two domains in one scalar Rquality = w c Q compute + w p Q grid. The grid quality term folds violations into an equivalent cost, where Qgrid = min(1, c/ (cθ + C LOL Viol)

Experimental Results

Experiments across eleven LLMs demonstrate the effectiveness of PowerAtlas under realistic physical operating conditions, with consistent feasibility and cost gains across three open-weight backbones. The gain holds across backbones, as violation falls by 84 to 98% and cost by up to 26× across the three backbones The alignment stage buys feasibility, while the warm start buys answerability

Online Deployment

The online deployment involves building an experimental electricity-computing network across three data centers in one province of China where the agent reads the live facility state and queue as the specification S The functional validation confirms that power-side schedules are checked against the model of Section 4.2 rather than actuated.

Conclusion

Neither stage alone suffices: the warm start buys answerability and the alignment stage buys feasibility, and they have to be applied in that order. The paper treats electricity-computing co-scheduling as one decision problem for an LLM agent<ref:2607.26710#pg9>

Acknowledgments

This work was supported by the Fundamental Research Funds for the Beijing University of Posts and Telecommunications under Grant 2025TSQY03. The experimental network of Section 6.6 was built with a provincial electric power company in China, which provided scoped access to its data-center facilities without disclosing security-sensitive operational data.

References

The paper lists numerous references supporting the work on LLMs for scheduling and power systems, including works on Unit Commitment Problem for Transmission System, Large Language Models can Schedule, and Security-Constrained Unit Commitment for Electricity Market. The references also include papers on LLMs for solving economic dispatch problem and Large Language Models as Optimizers. The paper includes specific references for the ECBench benchmark, and papers on data center workload management and spatial-temporal graph neural networks. The references also include foundational works on reinforcement learning for LLMs, and papers on power system optimization methods and DC Power Flow Problems Using Quantum and Hybrid Algorithms. The paper also cites works on LLMs for power scheduling, and papers addressing security risks in LLMs applied to power systems.

--- Page 1 ---

ECCS Task

The ECCS task is defined as generating a joint decision that simultaneously satisfies grid-security and task-service constraints given the grid physical model, the compute-task demands, and domain knowledge<ref:2607.26710#pg4> The input specification S comprises four parts (Figure 3), including grid physical constraints SC, task description ST, retrieved domain knowledge SK, and output-format contract SF<ref:2607.26710#pg4>

ECBench Dataset

PowerAtlas Framework

Improvements for AI systems

Bold header: Feasibility-Aware Policy Alignment (FA-GRPO)

The improved system can generate schedules that are jointly feasible and economically efficient by aligning the policy through Feasibility-Aware Group Relative Policy Optimization (FA-GRPO) which maximizes a reward function where the quality term is defined as Rquality = w c Q compute + w p Q grid.

Bold header: Model Agnostic Plug-in Architecture

The system can be deployed across different LLM backbones because PowerAtlas is a model-agnostic plug-in, validated on three backbones from three vendors and tokenizers, allowing it to match or beat the strongest frontier model on violation and cost.

Bold header: Verification via Deterministic Feasibility Program

The improved system ensures adherence to physical constraints by employing a deterministic verifier that parses the block, replays it through the hourly nodal DC feasibility LP, ensuring that a schedule is physically feasible and economically efficient at the same time.

Bold header: Benchmark for Coupled Optimization (ECBench)

The system's performance can be reliably evaluated against complex real-world scenarios because it utilizes ECBench, 2,000 standardized ECCS instances built from the RTS-GMLC system and real datacenter data, each paired with a Gurobi-optimal solution.

Bold header: Multi-Stage Training Strategy

The policy learns iteratively using two distinct stages: first, supervised fine-tuning to teach the output format and basic feasible decision patterns via LSFT, followed by RL alignment via FA-GRPO to optimize for the joint objective.