Researchers built a system that lets power grid operators ask questions about battery storage systems in plain English instead of writing database queries.
The framework, described in a new arXiv paper, sits between a large language model and a structured telemetry database tracking battery energy storage systems (BESS) on distribution networks. Operator questions get translated into validated SQL queries using predefined schema information and approved KPI views. The system pulls measurements like bus voltages, state of charge, active power, and reactive power, then checks them against engineering constraints for voltage limits and demand response tracking. The team validated it using hardware-in-the-loop co-simulation data from a battery-equipped distribution feeder running reactive power-based voltage control.
Grid operators already drown in telemetry as utilities bolt more batteries onto distribution networks for voltage regulation and demand response. The pitch here isn't a smarter grid - it's a faster way to interrogate the data the grid already produces, which matters more as BESS deployments scale past what manual dashboard-checking can handle. In testing, the framework caught repeated voltage violations and reactive power overshoot that a human sifting through raw logs might miss or catch late.
This is the LLM-as-translator pattern - natural language to SQL - showing up in a domain with real physical stakes, where a missed voltage violation isn't a bad query result but a grid reliability problem.