The rapid integration of power electronic devices such as grid-forming inverters into electrical grids has fundamentally transformed power system dynamics. These devices exhibit strong nonlinearity, multi-timescale interactions, and proprietary control algorithms that defy conventional modeling approaches. Traditional parameter identification methods assume known model structures, a limitation that becomes prohibitive when dealing with black-box control systems and incomplete prior knowledge about device behavior.
Researchers have introduced an innovative framework that leverages large language models (LLMs) and multi-agent collaborative intelligence to autonomously discover differential-algebraic dynamic models. The framework decomposes the discovery challenge into two complementary tasks: identifying the structure of differential equations governing state dynamics and determining algebraic closure relationships. Multiple specialized agents work in parallel, generating candidate equation structures while maintaining individual memory banks of successful candidates. A coordinator agent synthesizes findings and guides the collective search process, creating a feedback loop that refines model candidates through optimization and evaluation cycles.
The method's architecture integrates heterogeneous exploratory agents with unified measurement-data constraints, enabling joint recovery of system dynamics, algebraic constraints, and intermediate variables even when prior information is incomplete. This collaborative approach significantly outperforms single-agent LLM-based discovery and conventional symbolic regression techniques across multiple performance metrics.
Validation studies on synchronous generators achieved out-of-distribution mean absolute percentage error of just 0.19%, demonstrating exceptional generalization capability. For grid-forming inverters, the framework reduced discovery time by 25.7% compared to single-agent baselines while maintaining superior reconstruction accuracy and robustness to measurement noise. The framework's efficiency and accuracy improvements suggest substantial potential for accelerating the development of accurate models for novel grid technologies, ultimately supporting more effective grid planning, control design, and stability assessment in modern power systems with high renewable penetration.
