PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill-Based AI Agents for Power-System Studies

Pavel Etingov, Shuchismita Biswas

arXiv:2609.40272v1eess.SYcs.SY

TL;DR

Power-system studies involve many manual, tool-specific tasks that grow more demanding as operating conditions and uncertainties increase. This paper builds a skill-based multi-agent framework that connects agents to PSS®E through an MCP server and evaluates two implementation pathways. Both implementations executed the representative study tasks considered, while the evaluation was an initial proof of concept rather than a controlled benchmark.

  • Problem

    Power-system studies require many time-consuming, manual, tool-specific tasks dependent on expert knowledge, challenging scalability and timely decision support as scenarios and uncertainties increase.

    Source: §I · p. 1

  • Method

    The paper connects multi-agent workflows to PSS®E through a custom MCP server and reusable skills, evaluating implementations using the OpenAI Agents SDK and Claude Code across representative study procedures.

    Source: §I · p. 2, A. MCP Server for PSS®E Tool Integration · p. 2

  • Results

    Both implementations successfully executed the representative power-system study tasks considered, with qualitatively similar capability; the comparison was an initial proof of concept, not a controlled model or framework benchmark.

    Source: §IV · p. 4, §V · p. 5

  • Takeaways & Limitations

    The results indicate improved efficiency and repeatability for common analysis tasks; Claude Code was more practical when complex customization was unnecessary, while the SDK offered greater flexibility.

    Source: §V · p. 5

    The evaluation was an initial proof of concept without systematic repeated runs, and differing foundation models prevent a controlled framework or model comparison.

    Source: §IV · p. 4

Abstract

from arXiv · show

This paper describes a skill-based agentic framework for power-system studies using Model Context Protocol (MCP)-connected engineering tools. A custom MCP server was developed to expose Siemens PTI PSSE functions for power-flow analysis, dynamic simulation, result extraction, and model-validation workflows. Two implementation pathways built on a programmable OpenAI Agents software development kit (SDK) and a Claude Code command-line interface (CLI) were evaluated, both using reusable skills, subagents, MCP tools, data-repository connections, and local shell/Python execution. Both frontier-model-based implementations successfully executed representative study tasks. Success was evaluated based on task completion, output accuracy, and the need for human expert interventions. Results based on public datasets show that agentic systems can greatly accelerate power system dynamic simulation process for transmission planning studies leveraging industry-grade simulation platforms. This points toward a shift in transmission planning practice, where agentic systems could handle routine simulation setup and result extraction, allowing engineers to focus expert judgment on scenario design and interpretation rather than tool operation.

I. INTRODUCTION

Power-system studies combine numerous manual, tool-specific tasks, creating demands on expert time as operating conditions and uncertainties multiply. This paper proposes a skill-based multi-agent framework using MCP-connected engineering tools and evaluates it through representative workflows and two implementation pathways.

I. INTRODUCTION

  • Typical studies require many time-consuming, manual, tool-specific tasks, and the growing number of conditions and uncertainties challenges scalability and timely decisions.These workflows also depend on expert knowledge and significant human labor.Source: §I · p. 1
  • Prior work spans simulation with unseen tools, broader feedback-driven studies, distribution-grid analysis [10], solver-linked remediation [11], MCP and reusable skills,, and grid foundation models.A review also emphasizes safe, reliable, and accountable agent design.Source: §I · p. 1
  • The proposed architecture connects agents to Siemens PTI PSS®E, with an orchestrator delegating subtasks to specialized agents guided by reusable skills and controlled MCP access.Skills encode procedures, required inputs, tool-call sequences, validation checks, failure handling, and reporting templates.Source: §I · p. 2
  • The paper evaluates three representative procedures: power-flow and case analysis, dynamic simulation, and play-in model validation using synchrophasor or SCADA measurements.The two implementation pathways use the OpenAI Agents SDK and Claude Code.Source: §I · p. 2

II. AGENTIC IMPLEMENTATION AND DEPLOYMENT CONSIDERATIONS

The paper considers deployment choices for models and agent runtimes, including their trade-offs and potential grid-edge use. Locally deployed models can improve data control, while runtime options differ in flexibility and development effort.

II. AGENTIC IMPLEMENTATION AND DEPLOYMENT CONSIDERATIONS

  • Provider-hosted frontier models support rapid prototyping but may raise confidentiality and data-security concerns for utility cases involving planning, dynamic models, or sensitive measurements.The paper notes that on-premises or locally deployed models may be preferable for future production implementations.Source: §II · p. 2
  • Open-weight or locally deployable models support on-premises and edge use with better data control, offline operation, model-version stability, and provider independence, but may lag frontier models on complex engineering reasoning.Such models also require local compute infrastructure.Source: §II · p. 2
  • The OpenAI Agents SDK offers flexible orchestration, tool, guardrail, tracing, and MCP control but requires software-development effort; Claude Code needs less custom development.The passage contrasts programmable frameworks with skill-oriented environments.Source: §II · p. 2
  • Grid-edge agents could preprocess PMU or SCADA data locally and send compact summaries for engineering review; the studied model-validation use case is identified as suited to future edge deployment.The paper describes local preprocessing before transfer to centralized engineering systems.Source: §II · p. 2

A. MCP Server for PSS®E Tool Integration

A custom MCP server provides controlled, modular access to PSS®E through approved functions. It supports power-system study functions and custom play-in model validation, with skills guiding agents to handle problematic inputs explicitly.

A. MCP Server for PSS®E Tool Integration

  • The custom MCP server exposes approved PSS®E functions instead of allowing agents to manipulate the software, file system, or operating-system commands directly.Implemented in Python with FastMCP, it uses structured inputs and outputs and is designed to be modular, auditable, and reusable across agent runtimes.Source: A. MCP Server for PSS®E Tool Integration · p. 2
  • In the tested configuration, 21 PSS®E tools cover case management, power-flow solution, system-data queries, dynamic-simulation setup and execution, and result extraction.Figure 1 illustrates the server's functional blocks.Source: A. MCP Server for PSS®E Tool Integration · p. 2, Figure 1 · p. 3
  • The custom validation functions use play-in measurements to drive simulation and compare simulated active and reactive power responses with measured responses.Skills direct agents to report missing, corrupted, or inconsistent inputs rather than improvise unsupported values or actions.Source: A. MCP Server for PSS®E Tool Integration · pp. 2–3

B. Open AI Agents SDK Implementation

The OpenAI Agents SDK implementation routes user requests through an orchestration agent to specialized agents for study, validation, and analysis tasks. Skills guide their procedures, while sandboxed Python execution separates custom analysis and visualization from PSS®E execution.

B. Open AI Agents SDK Implementation

  • A chat interface passes study requests to an orchestration agent, which interprets the objective, identifies the workflow, and delegates to a specialized agent.Figure 2 presents the multi-agent architecture.Source: B. Open AI Agents SDK Implementation · p. 3, Figure 2 · p. 3
  • The orchestration agent maintains context and routes work to power-system study, model-validation, or code-and-visualization agents.The study agent handles power-flow and dynamic analysis; the validation agent handles play-in procedures.Source: B. Open AI Agents SDK Implementation · p. 3
  • Skills specify execution procedures, required inputs, tool-call sequences, validation checks, and expected outputs, while custom analysis and plotting run in a sandboxed Python environment separate from PSS®E.Source: B. Open AI Agents SDK Implementation · p. 3

C. Claude Code CLI Implementation

Claude Code coordinates study requests by selecting skills, delegating to a PSS®E subagent, using MCP tools, or running local post-processing. Reusable skills and integrated runtime capabilities reduce custom orchestration work and support possible event-driven workflows.

C. Claude Code CLI Implementation

  • Claude Code routes study requests among available skills, the PSS®E subagent, MCP tools, and local shell post-processing.The session coordinator chooses which capabilities to use for each request.Source: C. Claude Code CLI Implementation · p. 3
  • Reusable skills cover power-flow, dynamic simulation, play-in validation, and plotting, while the PSS®E subagent executes task-specific procedures through the shared MCP server.The subagent can open cases, solve power flow, load dynamics, run simulations, extract channels, and execute model-validation functions.Source: C. Claude Code CLI Implementation · p. 3
  • Claude Code requires less custom orchestration code; the setup also supports potential event-triggered validation and external disturbance-data screening.A tested API-connected skill accessed a PMU signature library for retrieving and screening disturbance records.Source: C. Claude Code CLI Implementation · p. 3

IV. RESULTS AND DISCUSSION

Both implementations completed the representative study tasks with qualitatively similar capabilities, but the evaluation was an initial proof of concept rather than a controlled benchmark. Claude Code required less custom development, and the PSS®E MCP server was registered and made available for task execution.

IV. RESULTS AND DISCUSSION

  • Both implementations executed the representative tasks with qualitatively similar capability, but results are an initial proof of concept without systematic repeated runs or a controlled model comparison.The implementations used GPT-5.5 and Claude Sonnet 4.6, respectively.Source: §IV · p. 4
  • Claude Code’s integrated runtime required less custom software development and made intermediate plans, tool calls, scripts, logs, outputs, and traces easier to inspect.The OpenAI Agents SDK implementation assembled and customized comparable capabilities through Python code.Source: §IV · p. 4
  • After registration, the PSS®E MCP server appeared as connected `psse-tools`, and Claude Code could discover its functions for task execution.Figure 4 presents the MCP-server setup in Claude Code.Source: A. Registration and tool discovery for… · p. 4, Figure 4 · p. 4

B. Power-Flow and Dynamic-Simulation Workflow

A single natural-language request specified a transmission-planning study combining power-flow setup, a timed bus fault, dynamic simulation, and CSV extraction. The workflow also tested skill-guided post-processing and visualization of simulation outputs.

B. Power-Flow and Dynamic-Simulation Workflow

C. Power-Plant Play-In Model-Validation Workflow

The play-in model-validation workflow applies domain-specific procedures to a customized power-plant validation task. The agent checked inputs, ran baseline and calibrated validations, extracted channels, and generated comparison plots.

C. Power-Plant Play-In Model-Validation Workflow

  • The play-in validation workflow coordinated measured event data, dynamic-simulation outputs, calibrated and uncalibrated model responses, and comparison plotting.Source: C. Power-Plant Play-In… · pp. 4–5
  • The agent checked for dynamic-data and play-in measurement files, ran baseline and calibrated validations, extracted output channels, and generated comparison plots.Figure 7 presents the execution trace and output.Source: C. Power-Plant Play-In… · p. 5, Figure 7 · p. 5
  • The example supports specialized engineering workflows when their procedures are encoded in skills and exposed through MCP-connected tools.Source: C. Power-Plant Play-In… · p. 5

V. CONCLUSIONS

The framework connected AI agents to selected PSS®E functions, and both implementations completed representative study tasks. Results indicate efficiency and repeatability gains, while future work targets broader functions and secure local deployment.

V. CONCLUSIONS

  • A custom MCP server exposed selected PSS®E functions for power-flow analysis, dynamic simulation, result extraction, plotting, and model validation.The two evaluated pathways used the OpenAI Agents SDK and Claude Code CLI.Source: §V · p. 5
  • Both implementations completed representative tasks; agentic approaches improved efficiency and repeatability, with Claude Code more practical when complex customization was unnecessary.The OpenAI Agents SDK offered greater flexibility for customized orchestration and production-style applications.Source: §V · p. 5
  • Future work will expand supported analyses and evaluate on-premises and grid-edge deployment with locally hosted models and secure execution environments.Planned capabilities include contingency analysis, batch dynamic simulations, parameter calibration, oscillation analysis, and automated reporting.Source: §V · p. 5
Loading 2609.40272v1…