Intent-Driven LLM Agent Framework for Power System Simulation

This article introduces an intent-driven agent framework based on large language models (LLMs), which aims to automate power system simulation workflows, decompose complex simulation tasks into executable steps, and improve the efficiency and accessibility of power system analysis. This framework addresses the high barrier to entry of traditional simulation tools, allowing non-professional users to use them conveniently and having broad application prospects.

Complexity Challenges of Power System Simulation

Power system simulation involves various tasks such as power flow calculation and short-circuit analysis, relying on complex toolchains like PSS/E and PSCAD. It requires users to have deep professional knowledge and operational skills. Manual steps are tedious, and data transfer across tools is difficult, making it hard for non-professional users to use.

New Opportunities for LLMs

LLMs can serve as intelligent interfaces to understand natural language intent, generate scripts/call APIs, explain results, and refine tasks through multi-round interactions. However, directly connecting to simulation tools faces challenges such as complex interfaces and insufficient domain knowledge support, requiring systematic framework organization capabilities.

The framework adopts the intent-driven concept: users describe tasks in natural language, and the system automatically completes understanding, planning, execution, and result return. The core components include:

1. Intent Understanding Module: Extracts key information such as simulation type, target power grid, scenario, and output requirements;
2. Task Planning Module: Decomposes high-level intent into a sequence of subtasks (e.g., loading models, setting parameters, executing simulations, etc.);
3. Tool Calling Module: Interacts with simulation software through function calls or code generation;
4. Result Processing Module: Analyzes raw data and generates user-friendly reports;
5. Interaction Management Module: Clarifies intent through multi-round dialogues and handles simulation failures.

Multi-Agent Collaboration

The main agent coordinates tasks, while professional agents handle specific domain work such as data preprocessing and power flow calculation.

Domain Knowledge Integration

Integrates power domain knowledge through LLM context constraints, Retrieval-Augmented Generation (RAG), and built-in simulation templates.

Safety and Verification Mechanisms

Includes multi-layer guarantees such as parameter range checks, result consistency verification, operation confirmation, and audit logs.

Scalability Design

The modular architecture supports the integration of new simulation tools and data sources, and defines standard interfaces to simplify integration.

This framework has wide applications in the power system field:

• Teaching and Training: Reduces learning barriers, allowing students to focus on understanding principles;
• Rapid Prototype Verification: Accelerates research iteration without complex scripts;
• Decision Support: Provides convenient analysis for managers;
• Automated Report Generation: Executes tasks in batches and generates reports;
• Intelligent Fault Diagnosis: Assists operation and maintenance personnel in analyzing and handling faults.

Compared to general frameworks like AutoGPT and LangChain, this framework is optimized for power simulation:

• Domain-Specific Toolset: Encapsulates professional interfaces of power simulation software;
• Physical Constraint Awareness: Understands the physical laws of power systems;
• Result Credibility Evaluation: Designs verification mechanisms for simulation results;
• Professional Terminology Understanding: Optimizes the processing of power domain terminology.

Vertical domain customization is the key path for LLMs to move from general to professional use.

This research is an important exploration of LLMs in the industrial software field, and similar ideas can be extended to mechanical finite element analysis, fluid dynamics simulation, BIM, and other fields. The framework can reduce the threshold for using complex industrial software, improve engineering efficiency, and promote knowledge dissemination. It also reveals common challenges: correctness of generated code, handling of complex constraints, integration of legacy systems, etc., laying the foundation for industrial AI applications.