PaperlayerView exact revision on arXiv

Source-linked AI summary

Grid-Agent: An LLM-Powered Multi-Agent System for Power Grid Control

Yan Zhang, Ahmad Mohammad Saber, Amr Youssef, Deepa Kundur

arXiv:2508.05702v3cs.MAcs.AIcs.SYeess.SY

TL;DR

Modern grids’ variable operating conditions challenge traditional optimization, while LLM-based multi-agent control of power grids remains largely unexplored. Grid-Agent combines LLM reasoning, numerical solvers, adaptive network representation, and validation mechanisms to resolve violations. Across tested models, reported success rates range from 44% to 100%, with future scalability to transmission-level networks still to be examined.

  • Problem

    Traditional optimization can struggle with dynamic grid conditions, while LLM-based multi-agent applications to mission-critical power grids remain largely unexplored.

    Source: §I · p. 1

  • Method

    Grid-Agent combines LLM semantic reasoning and numerical solvers in a multi-agent framework with adaptive network representation and sandboxed validation.

    Source: §I · p. 1, C. Adaptive Multi-Scale Network… · p. 4, D. Safety, Validation, and Rollback… · p. 5

  • Results

    Success rates were 100% for gemini-2.5-flash and gpt-4.1-mini, while gemini-2.5-flash-lite and gpt-4.1-nano achieved 63% and 44%, respectively.

    Source: D. Experimental Results and Analysis · p. 6

  • Takeaways & Limitations

    Grid-Agent’s coordinated reasoning can identify strategic actions, such as a single switch operation resolving multiple downstream issues, rather than addressing violations sequentially.

    Source: §VI · p. 9

    Scalability to transmission networks with thousands of buses remains future work and may require advances in hierarchical representation and distributed multi-agent planning.

    Source: §VI · p. 9

Abstract

from arXiv · show

Modern power grids face unprecedented complexity from Distributed Energy Resources (DERs), Electric Vehicles (EVs), and extreme weather, while also being increasingly exposed to cyberattacks that can trigger grid violations. This paper introduces Grid-Agent, an autonomous AI-driven framework that leverages Large Language Models (LLMs) within a multi-agent system to detect and remediate violations. Grid-Agent integrates semantic reasoning with numerical precision through modular agents: a planning agent generates coordinated action sequences using power flow solvers, while a validation agent ensures stability and safety through sandboxed execution with rollback mechanisms. To enhance scalability, the framework employs an adaptive multi-scale network representation that dynamically adjusts encoding schemes based on system size and complexity. Violation resolution is achieved through optimizing switch configurations, battery deployment, and load curtailment. Our experiments on IEEE and CIGRE benchmark networks, including the IEEE 69-bus, CIGRE MV, IEEE 30-bus test systems, demonstrate superior mitigation performance, highlighting Grid-Agent's suitability for modern smart grids requiring rapid, adaptive response.

I. INTRODUCTION

Modern grids’ variable conditions challenge traditional optimization and leave LLM-based multi-agent control of critical infrastructure largely unexplored. Grid-Agent explores combining semantic reasoning with numerical analysis for grid-violation remediation and reports tests on IEEE and CIGRE benchmarks.

A. Power Systems Optimization

  • Traditional optimization methods can struggle to generalize, interpret context, or reason semantically under dynamic grid conditions.Classical approaches also face convergence challenges on complex mixed-integer, non-convex, nonlinear, unbalanced AC optimal power flow problems and difficulty incorporating operator knowledge or unforeseen events.Source: §I · p. 1, A. Power Systems Optimization · p. 2

I. INTRODUCTION

  • LLM-based multi-agent applications to mission-critical power grids remain largely unexplored, despite opportunities to combine natural-language reasoning with numerical precision.The paper frames Grid-Agent as an exploration of agentic reasoning and coordinated planning, not a finalized production-ready solution or a pursuit of formal control-action optimality.Source: §I · p. 1
  • Grid-Agent combines LLM-based semantic reasoning with numerical solvers to detect and remediate violations through coordinated optimization of topology, batteries, and demand response.Its adaptive multi-scale representation adjusts encoding granularity, and a continuous-learning pipeline generates training data and fine-tunes LLM components.Source: §I · p. 1
  • Benchmark tests on IEEE and CIGRE systems report superior violation mitigation, scalability across network sizes and complexities, and robustness to simultaneous violations, including cyberattack-induced contingencies.The tested systems include IEEE 69-bus, IEEE 30-bus, and CIGRE MV networks.Source: §I · p. 1

B. Large Language Models in Engineering Applications

Prior LLM energy applications mainly address analysis and prediction rather than direct control, while multi-agent systems face power-grid safety and real-time constraints. The paper introduces grid-violation resolution as a problem involving operational limits and physical isolation.

B. Large Language Models in Engineering Applications

  • Energy-sector LLM applications have focused mainly on analytics, trading, and demand-response optimization, leaving active real-time control and optimization in critical infrastructure as a gap.The passage contrasts these uses with Grid-Agent’s focus on direct system control.Source: B. Large Language Models in… · p. 2

C. Multi-Agent Systems with LLMs

  • LLM multi-agent systems use reasoning within structured workflows and state management, but power-grid deployment adds safety, real-time, and explainability challenges.Related examples include software engineering, automated scientific research, and collaborative problem-solving.Source: C. Multi-Agent Systems with LLMs · p. 2

III. GRID-AGENT: PRELIMINARIES

  • The preliminaries define grid-violation resolution through the network’s physical laws and the discrete and continuous control actions available to the agent.The supplied section introduction frames the problem before specifying violation types and limits.Source: §III · p. 2

A. Power Grid Violation Resolution Problem

  • The paper identifies voltage-limit breaches, line apparent-power or current ratings being exceeded, and buses electrically isolated from the main grid as violation categories.An isolated bus can result in complete loss of power supply.Source: A. Power Grid Violation Resolution… · p. 2

B. Power Flow Constraints

Power-flow constraints require bus injections to satisfy nonlinear balance equations that connect voltage magnitudes and angles across the network, so proposed actions yield physically possible states.

B. Power Flow Constraints

  • Net active and reactive power injections at each bus must be balanced under the fundamental power-flow equations.These equations encode the physical laws governing the network.Source: B. Power Flow Constraints · p. 2
  • The nonlinear power-flow equations connect bus voltages and angles across the entire network.The equations apply at each bus in the network.Source: B. Power Flow Constraints · p. 2
  • The equations use conductance and susceptance terms from the bus admittance matrix to ensure proposed control actions yield physically possible grid states.The prose identifies Gij and Bij as the conductance and susceptance terms, respectively.Source: B. Power Flow Constraints · p. 2

C. Action Space Constraints

Grid-Agent’s control actions span switch operations, load curtailment, and battery placement and dispatch, with battery operation coupled to installation and equipment limits.

C. Action Space Constraints

  • Switch operations are binary decisions to open or close designated switchable lines.The switchable-line set is a subset of the network’s lines.Source: C. Action Space Constraints · p. 3
  • Load curtailment continuously reduces active demand at curtailable-load buses using a curtailment fraction.The fraction determines the resulting load value.Source: C. Action Space Constraints · p. 3
  • Battery actions jointly choose installation buses and active and reactive dispatch, constrained by battery availability and power ratings.Dispatch is permitted only where a battery is installed, and the formulation keeps allocation and dispatch within physical limits.Source: C. Action Space Constraints · p. 3

D. Large Language Models and Agentic AI

Grid violation resolution requires reasoning across complex scenarios and heterogeneous information, capabilities the section attributes to LLMs and agentic AI. LLMs combine semantic interpretation, contextual reasoning, and tool-based numerical validation.

D. Large Language Models and Agentic AI

1) LLMs as Reasoning Engines for Grid Management:

  • LLMs can interpret grid numbers alongside topology descriptions, operational logs, and maintenance reports, supporting a more holistic understanding of grid states.They can also incorporate domain knowledge and constraints such as prioritizing critical loads or avoiding switching near vulnerable points.Source: 1) LLMs as Reasoning Engines for Grid… · p. 3
  • Tool-augmented LLMs delegate complex calculations to validated power flow solvers, grounding high-level reasoning in grid physics and operational constraints.Source: 1) LLMs as Reasoning Engines for Grid… · p. 3

2) Multi-Agent Systems for Decomposed Problem Solving:

Multi-agent systems divide violation resolution into specialized tasks and coordinate agents through an explicit workflow. Grid-Agent applies this approach in a state-driven process that combines LLM semantic reasoning with validated power-system simulations.

2) Multi-Agent Systems for Decomposed Problem Solving:

  • Specialized agents handle tasks such as topology analysis, action planning, solution validation, and result summarization, with iterative Planner–Validator feedback refining strategies.The modular interactions are described as supporting scalability, interpretability, and auditable reasoning across grids of varying complexity.Source: 2) Multi-Agent Systems for Decomposed… · p. 3
  • The workflow W comprises agents A, their flow relationships F, and conditional branching logic C; its state records the network configuration, violations, actions, and intermediate results.Source: 2) Multi-Agent Systems for Decomposed… · p. 3

IV. GRID-AGENT: PROPOSED SYSTEM ARCHITECTURE AND METHODOLOGY

  • Grid-Agent uses a state-driven workflow in which specialized agents plan, execute, and validate actions, combining LLM semantic reasoning with validated power-system simulations.Source: §IV · pp. 3–4

A. State-Driven Multi-Agent Workflow

Grid-Agent cycles through five specialized agents to identify violations, plan and execute control actions, validate outcomes, and summarize successful resolutions. Its planner uses a structured LLM prompt and an iterative action-evaluation loop.

A. State-Driven Multi-Agent Workflow

  • The five-agent workflow parses grid conditions, plans multi-step actions, executes them through a solver on a sandbox copy, validates for resolved violations and new instability, and rolls back failed plans.After successful resolution, the Summarizer creates a human-readable explanation and records the process for continuous-learning datasets.Source: A. State-Driven Multi-Agent Workflow · p. 4

B. LLM-Guided Coordinated Action Planning

  • The Planner receives network state, active violations, available actions, and operational constraints, then produces machine-parsable tool calls for topology changes, battery deployment or dispatch, and load curtailment.The prompt frames the LLM as an expert operator and guides it with a prioritized action policy.Source: B. LLM-Guided Coordinated Action… · p. 4
  • The planning loop initializes a sandbox, analyzes violations, generates a network representation and LLM plan, executes actions, rechecks violations, evaluates effectiveness, and updates knowledge until resolution or the iteration limit.Source: B. LLM-Guided Coordinated Action… · p. 4

C. Adaptive Multi-Scale Network Representation

Grid-Agent adapts its network representation to system size and complexity, pairing that scalability approach with validation and rollback safeguards. Successful resolutions are documented as structured data for potential LLM fine-tuning.

C. Adaptive Multi-Scale Network Representation

  • For smaller networks, Grid-Agent serializes every component; for larger networks or clustered violations, it abstracts healthy sections into a semantic graph focused on violated components and nearby controllable assets.The approach is intended to scale from microgrids to large distribution networks without overwhelming the LLM with excessive data.Source: C. Adaptive Multi-Scale Network… · pp. 4–5

D. Safety, Validation, and Rollback Mechanisms

  • Proactive solver checks, post-execution assessment for resolved and newly introduced violations, and rollback of plans that fail to improve the grid state form a layered safety strategy.Source: D. Safety, Validation, and Rollback… · p. 5

E. Continuous Learning Framework

  • After successful resolution, the system records the initial state and violations, successful action sequence, and explanation in a structured training instance for possible LLM fine-tuning.Source: E. Continuous Learning Framework · p. 5

V. EXPERIMENTAL SETUP AND VALIDATION

The evaluation compares Grid-Agent across standardized IEEE and CIGRE test scenarios, six Planner Agent LLMs, and metrics covering resolution, efficiency, quality, and runtime.

A. Test Network Configuration

  • IEEE 30-bus, CIGRE MV 15-bus, and IEEE 69-bus scenarios include concurrent violations and disconnections, with automated generation supporting repeatable tests.The cases range from 2 voltage and 3 thermal violations in Case30 Light to 29 violations in IEEE 69 Medium Loads.Source: A. Test Network Configuration · pp. 5–6, Table I · p. 6

B. LLM Test Suite Configuration

  • Six LLMs were integrated into the same Planner Agent role and compared, with Gemini thinking budgets minimized to prioritize inference speed.The models were gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gpt-4.1, gpt-4.1-mini, and gpt-4.1-nano.Source: B. LLM Test Suite Configuration · p. 6, Table II · p. 6

C. Performance Metrics

  • Metrics assess full scenario resolution, violations resolved per action, iterations to resolution, solution quality, and wall-clock runtime.Solution quality favors fewer control actions and avoiding disruptive measures such as load curtailment.Source: C. Performance Metrics · p. 6

D. Experimental Results and Analysis

Grid-Agent’s performance varied by LLM: four models achieved 100% success, while the two smaller models struggled. Gemini-2.5-flash combined high success with fast runtimes, and action efficiency rose on the larger radial IEEE 69-bus network.

D. Experimental Results and Analysis

VI. DISCUSSION

Grid-Agent combines LLM reasoning with numerical optimization and coordinated control actions, with experiments on IEEE and CIGRE networks supporting its effectiveness. The discussion highlights strategic actions and interpretable explanations while identifying transmission-scale testing and formal optimality analysis as future work.

VI. DISCUSSION

  • A strategic switch operation can resolve multiple downstream issues at once, contrasting with myopic algorithms that address violations sequentially.The discussion links this coordinated approach to high action efficiency and holistic reasoning in interconnected violations.Source: §VI · p. 9
  • Step-by-step explanations provide control-action rationale, which the discussion identifies as important for operator confidence and regulatory compliance.The paper presents this as a way LLM-based systems may enhance grid operations through human-interpretable reasoning.Source: §VI · pp. 8–9
  • Adaptive multi-scale representation adjusts LLM input detail to network size and complexity, addressing limited context windows while preserving solution quality.Future work should examine transmission networks with thousands of buses, where hierarchical representation and distributed coordination may be needed.Source: §VI · p. 9
  • Formal solution optimality remains unassessed; planned benchmarking against mixed-integer nonlinear OPF solvers would quantify trade-offs with heuristic, rapid, interpretable decisions.The discussion frames this comparison as a way to establish where agentic approaches excel and traditional optimizers remain indispensable.Source: §VI · p. 9

VII. CONCLUSION

  • Grid-Agent integrates LLM reasoning with power-system optimization, and IEEE and CIGRE tests support its effectiveness across diverse violation scenarios, especially complex multi-violation cases.The conclusion also reports sandboxed execution and automated rollbacks to validate proposed solutions before they affect the system.Source: §VII · p. 9
Loading 2508.05702v3…