Data-driven digital twin-assisted energy storage scheduling for electric vehicle charging stations under demand uncertainty: explainable forecasting and multi-objective dispatch

Objectives Public electric-vehicle charging sites are often planned from transaction records without charger telemetry or an installed battery energy storage system (BESS). This study evaluates whether such data can support pre-deployment storage scheduling under uncertain demand. Methods We analyzed 441,077 charging transactions from 13 sites in Jiaxing, China. After prespecified screening, 348,404 sessions (6.754 GWh) were reconstructed into 30-min station demand. LightGBM forecasting was evaluated on an untouched October–December 2021 test period across 12 stations. Additional analyses included station-day cluster bootstrap, station-wise validation, an independently trained median-quantile model, station-conditional split-conformal calibration, and energy-conserving reconstruction sensitivity. An offline decision twin evaluated a 79.15 kW/158.31 kWh virtual BESS at Bus Station. Results LightGBM achieved RMSE 29.37 kW and R 2 = 0.647, compared with 29.86 kW for the seven-day same-slot baseline. A 10,000-resample cluster bootstrap estimated an RMSE improvement of 0.493 kW (95% CI 0.287–0.701), equivalent to 1.65% (0.96%–2.34%). LightGBM improved RMSE at 10 of 12 stations. The original q05–q95 interval covered 79.56% of test observations; split-conformal calibration increased coverage to 89.03% with mean width 76.86 kW. The independently trained median model was highly correlated with, but not equivalent to, the point model (r = 0.992; RMSE 30.17 vs. 29.37 kW). Reconstruction sensitivity changed cost and peak reductions by at most 0.03 and 0.08 percentage points, respectively. Uncertainty-aware dispatch reduced modeled expected cost by 3.31% and q95 peak demand by 3.14%. Conclusion The results support a modest but stable forecasting gain, reveal station heterogeneity, and indicate limited sensitivity to the tested reconstruction assumptions. All storage benefits remain conditional simulation outcomes from an offline pre-deployment decision model.