LLM-Sim: An Intelligent Agent Framework for Natural Language-Driven Power System Simulation

LLM-Sim is an open-source project developed by samimk, which innovatively integrates large language models (LLMs) with the high-performance power system simulation tool ExaGO, enabling a "conversational" grid analysis paradigm. Users only need to describe analysis objectives in natural language, and the system automatically completes the full closed loop of parameter adjustment, simulation execution, and result interpretation, bringing new possibilities to power system analysis.

Power system analysis has always been highly specialized, requiring engineers to master complex models, algorithms, and tools. However, with the large-scale integration of renewable energy and the complexity of power markets, traditional manual analysis can no longer meet the needs of rapid decision-making. The emergence of LLM-Sim aims to address this pain point by combining the reasoning capabilities of LLMs with ExaGO's optimization platform.

The core of LLM-Sim is an iterative agent loop that simulates the thinking process of human engineers: starting with parsing grid cases in MATPOWER format, it first runs a baseline simulation to obtain the initial state; the LLM receives prompts of user objectives, grid overview, and simulation results, then decides the next action (modify parameters and re-simulate, request data analysis, or complete the task); the key "search log" mechanism records each iteration result, provides historical context, helps the model avoid invalid paths, and gradually approaches the optimal solution.

LLM-Sim relies on the ExaGO platform developed by the Oak Ridge National Laboratory (ORNL) in the United States. This platform is designed for parallel/distributed architectures to solve large-scale grid optimization problems. It supports scenarios such as AC optimal power flow (OPFLOW), multi-period/safety-constrained/stochastic optimal power flow, and can work with solvers like Ipopt and HiOp, running on CPU or GPU acceleration. LLM-Sim seamlessly combines the "soft reasoning" of LLMs with the "hard computing" of ExaGO, balancing interaction convenience and result accuracy.

LLM-Sim covers multiple power system analysis scenarios: boundary exploration (e.g., finding the maximum load scaling factor before the system becomes infeasible); scenario analysis (e.g., simulating the impact of generator tripping); optimization problems (e.g., minimizing power generation costs under voltage constraints); and various analysis tasks such as identifying congested lines, evaluating voltage distribution, and generating operation reports.

LLM-Sim supports real-time interaction, allowing users to inject new instructions during simulation (without stopping or restarting). Intervention methods include: enhancement mode (adding constraints/preferences, e.g., "focus on bus in area 3"); replacement mode (completely overwriting the current objective). It also supports pausing, resuming, or terminating the search, making it suitable for both fully automated batch processing and complex analyses requiring manual intervention.

LLM-Sim is developed in Python and can be installed via pip, providing a Shell script launcher (requiring configuration files, case files, and iteration limits). It supports LLM backends such as Anthropic Claude and OpenAI GPT, with configurations managed via YAML files (backend selection, application type, log level, etc.). The interface includes a concise command line (supporting dry run and quiet mode) and a Streamlit-based graphical interface (for real-time monitoring and interaction).

LLM-Sim demonstrates an important application direction of AI in the professional engineering field, transforming human intentions into precise technical operations. For engineers: it lowers the threshold for complex analysis and accelerates decision-making (especially in scheduling scenarios); for researchers: it provides a scalable platform to explore human-machine collaborative energy optimization. With the digital transformation of power grids, such intelligent agent tools will bridge professional knowledge and general AI capabilities, enhancing system understandability and controllability.