When Do EV-Charging Load Forecasters and Controllers Transfer? A Cross-Dataset Study of Interval Calibration, Negative Transfer, and Forecast-Informed Control

Deep forecasting and learning-based control are widely proposed for electric-vehicle (EV) charging hubs, yet theliterature almost universally reports accuracy on a single dataset and rarely asks whether a model trained on onesite, or on synthetic data, retains skill elsewhere. We study transferability directly. A quantile long short-termmemory (LSTM) forecaster with conformally calibrated 90% prediction intervals is trained on a large UnitedKingdom multi-provider dataset (≈155k sessions, 2017–2022) and evaluated against the Caltech ACN-Data huband a SimPy-generated synthetic stream under four transfer directions and a zero-shot/fine-tune protocol. Threefindings emerge. First, point-forecast skill does not transfer: a model with native test R²=0.91 (UK) and R²=0.85(ACN) collapses to R²=0.06 synthetic→real and R²=−4.19 zero-shot across hubs, and even hub-specific LSTMsfail to beat a naive persistence baseline (Diebold–Mariano p=0.138; ACN persistence R²=0.863 vs native LSTMR²=0.852). Second, interval calibration is the property that survives: split-conformal recalibration restoresempirical coverage to near-nominal on the native distribution, and calibrated synthetic data lifts crossdistribution R² by up to +2.34. Third, model-predictive control built on forecasts generalises where pointaccuracy does not: cutting curtailment 25–55% versus a tuned PID controller, a single fixed MPC configurationholds the hard capacity constraint (at most one overload) across four hubs spanning 12–188 kWh even as theforecaster's cross-site skill ranges from R²=0.91 down to R²=−3.96 — the controller's safety guarantee transfersalthough the forecaster does not. We argue that EV-charging studies should report calibrated intervals and crosssite transfer as first-class results, and we release all metrics and code to support replication