LLMs and Agentic AI Systems for Smart Grids: A Tutorial on Architectures and Applications
Daniela Rojas, Abdulwahab Albassam, Aidan G. Leung, Jett Ngo, Ryan Luo, Peter R. Quawas, Junpyung Kim, Kangkai Liang, Mansi Nanavati, Jonathan Mai, Meng-Chi Tsai, Yun-Tong Tsai, Yize Chen, Yuanyuan Shi
Read on arXiv →
Key claim
Solver-grounded design enhances LLM reliability in power systems.
In plain English
Imagine you're building a smart grid management system that needs to forecast energy demand, optimize resource allocation, and ensure system stability. The challenge lies in ensuring that the AI can not only generate plausible solutions but also verify their feasibility against real-world constraints. Current approaches often lead to outputs that sound good but can be physically impossible, which is a problem known as hallucination. This can result in significant operational failures, especially in critical systems like power grids where incorrect decisions can have serious consequences. This is what's called a reliability issue in AI systems.
To address these challenges, the authors propose a solver-grounded design principle, which ensures that any numerical result reported by the AI comes from a trusted tool and has undergone explicit verification. This approach integrates LLMs with reliable solvers, allowing for a more robust orchestration of tasks. They present four case studies in power systems, demonstrating how this principle can significantly improve outcomes, such as reducing unmet energy demands and effectively diagnosing system contingencies. By establishing a clear evaluation framework that assesses task utility, correctness, and safety, the work provides a structured way to build and evaluate AI systems that can be trusted in real-world applications.
The paper introduces a solver-grounded design principle that enhances LLMs' reliability in technical domains.
The evaluation framework and case studies provide solid evidence of the proposed approach's effectiveness.
Deep reliability assessment
The methodology supports the use of trusted tools for numerical results, ensuring outputs are physically feasible, but it may overclaim the ability of LLMs to handle all aspects of smart grid tasks without further empirical validation.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
Figure 8 illustrates the wind forecasting trajectory using the APBF method, showing the agent's process of binning history, retrieving weather forecasts, constructing a physics-informed prompt, and validating the JSON schema.