Arrival Time Occupancy Forecasting for Electric Vehicle Charging Recommendations Using Hybrid Machine Learning and Queueing Models

A highway driver choosing where to charge needs to know whether a charger will be free when the car arrives, not whether it is free when the trip starts. Amplyt is a charging advisor built around that need. For each station it forecasts connector occupancy at the predicted arrival time and passes the forecast through a finite-capacity queueing model and an isotonic calibration layer. A data audit early in development showed that the original wait-time target could be rebuilt from availability recorded at the same moment, so the target was redefined and the affected workbook withdrawn. The cleaned corpus holds 463 observations of sixteen stations, and the gradient-boosted model was trained on 724 connector-level trials from thirteen of them. In date-blocked rolling-origin evaluation it scored a Brier of 0.1965 against 0.2183 for a station-hour baseline, but most of the gain comes from quiet stations: at the five busiest the improvement is 1.9 percent and statistically indistinguishable, and leave-one-station-out tests are worse at unseen stations. The end-to-end wait pipeline does not beat an always-zero baseline, so forecast minutes are withheld. Arrival-time availability can be shown honestly as bands with confidence tiers, but the data does not yet support wide deployment. Key Words: Electric vehicle charging, occupancy forecasting, queueing model, gradient-boosted trees, isotonic calibration, data leakage.