The increasing adoption of Electric Vehicles (EVs) is introducing new sources of uncertainty into power system operation. Accurate forecasts of EV charging sessions (e.g., start time, duration, and energy consumption) are becoming a prerequisite to anticipate peak loads, manage network constraints at the distribution level, and coordinate charging with renewable generation and energy storage solutions, enabling EV demand to be operated as a flexibility resource rather than a passive load.

Large Language Models (LLMs) have shown competitive performance on time-series tasks formulated as text, but it is unclear how they compare with established forecasting methods on real-world EV charging data. As part of an ongoing research project on data-driven EV charging forecasting, this study benchmarks LLM-based forecasting approaches against naive (e.g., moving averages and median) and classical machine-learning baselines (e.g., linear regression and gradient boosting machines), at two levels: forecasting the next session(s) of an individual user, and forecasting the next session(s) at an individual charger. The analysis is conducted on real-world charging data from multiple chargers (Electric Vehicle Supply Equipment, EVSE) operated in Porto, Portugal, comprising 2,933 sessions from multiple users across 12 chargers over 2 years.

Results are target dependent. For EV charging sessions start-time forecasting, LLM-based approaches outperform all baselines, with nearly 40% reduction in Mean Absolute Error (MAE) over the strongest baseline, both per-user (from 31.92 to 19.06 hours) and per-charger (from 27.61 to 14.39 hours). For energy transferred and session duration, however, LLM-based approaches are on par with simpler methods such as moving averages and linear regression. This suggests LLMs capture temporal regularities in charging behaviour but not magnitude, motivating hybrid approaches that combine LLMs for timing with classical regressors for energy and duration.