From Natural Language to Optimal Schedules: An LLM Interface for EV Charging

From Natural Language to Optimal Schedules: An LLM Interface for EV Charging XUE Sihui, First-year Ph.D. Student KODAIRA Daisuke, Assistant Professor Graduate School of Science of Technology Smart Grid Lab

Background Problem Framework Case study Conclusion “I want my EV to be charged by 7 a.m.” Incomplete config Input - Arrival time: Unknow - Departure time: 7:00 - Target SoC: Full or 90% Grid - Max power: Unknow ··· Electric vehicle Optimizer - V2G: Unknow Invalid input • EV number increases, concentrated charging stresses the grid: load peaks and price volatility. • Optimization model (e.g., linear program) can schedule charging, but only with complete, structured inputs. • Natural language intent is incomplete and ambiguous, so it cannot be executed by the optimizer directly. -2-

Background Problem Framework Case study Conclusion • Large language model (LLM) applications in energy systems: mostly offline or advisory Offline analysis and reporting Forecasting tasks Operational support and scheduling recommendations Building and home energy management Grid operations and fault diagnosis Result explanation and conversational interfaces -3-

Background Problem Framework Case study Conclusion • Problem statement “I want my EV to be charged by 7 a.m.” Incomplete config Input - Arrival time: Unknow - Departure time: 7:00 - Target SoC: Full or 90% - Max power: Unknow Optimizer - V2G: Unknow Invalid input ➢ Challenge 1: Natural language requests are incomplete and ambiguous, important parameters may be missing. ➢ Challenge 2: Optimizers require validated, structured inputs. -4-

Background Problem Framework Case study Conclusion • From natural language to optimal schedules I want my EV to be fully charged by 7 a.m., preferably avoiding high-price periods, and the charging power should stay below 7 kW, save my money. User ② Extracted config - Arrival: 18:00 - Departure: 7:00 ① Natural-language - Target SoC: 100% request - Max power: 7 kW ⑥ Explanation LLM agent ③ Model config - V2G: on ⑤ Results ④ Schedule EV Optimizer (LP) -5-

Background Problem Framework Case study Conclusion • Case study setup (LP model, ChatGPT-5-mini model) Time-of-Use price Battery capacity 60 kWh Timestep 1 hour Battery SoC between 20% to 90% Battery efficiency 0.95 • Validated LLM-extracted parameters enable effective LP scheduling. (a) Charging power profiles of the two strategies (b) SoC trajectories of the two strategies -6-

Background Problem Framework Case study Conclusion • Experimental process of the case study Agent-User interaction User: I will arrive home around 6 p.m. and leave at 8 a.m. tomorrow. Please charge my EV from about 20% to at least 80% by departure, avoid high-price periods as much as possible. Extracted configuration - Initial SoC: 0.2, - Target SoC: 0.8, - Start Hour: 18, - End Hour: 8, - V2G: false Explanation Agent: Schedule generated. 64.9% (¥28.00) saved by shifting charging to the 23:00–06:00 valley. Target 80% SoC ensured by 08:00. -7-

Background Problem Framework Case study Conclusion • Summary ➢ We proposes an LLM-based framework that converts EV charging requests in natural-language into a structured parameters and drives the LP optimizer effectively. ➢ It makes the optimization model easier for non-technical people to use. • Future work ➢ User side: better handling of more complex user requests ➢ Grid side: show simple indicators of grid impact ➢ Compare different optimization model for charging based on our framework. -8-

Thank you all for listening. XUE Sihui xue.sihui.tbk_gb@u.tsukuba.ac.jp KODAIRA Daisuke kodaira.daisuke.gf@u.tsukuba.ac.jp Graduate School of Science of Technology Smart Grid Lab