Replication package for: Uncertainty-Guided Adaptive Metamodeling for Discrete-Event Simulation of Urban Electric-Vehicle Charging Networks

Planning public electric-vehicle (EV) charging networks requires knowing how demand, charging rates, tariffs, and congestion-induced derating affect waiting, abandonment, throughput, and revenue, but discrete-event simulators that capture driver patience and station switching are too costly to run for every planning scenario. This study asks how few simulation runs are needed to learn such a network's behavior, using a spatial urban charging-network simulator implemented in AnyLogic and replicated in Python, with five responses: waiting time, total revenue, throughput, utilization, and driver dissatisfaction. Starting from the same 30-run space-filling design, with every point averaged over five replications, two 15-run augmentation strategies are compared: residual-guided linear refinement and multi-output Gaussian-process (GP) uncertainty sampling. At 45 runs, the GP branch achieved a mean normalized leave-one-out cross-validation (LOOCV) RMSE of 0.168 versus 0.485 (65.4% lower). Under a common GP evaluator, LOOCV favored the residual-guided design (0.107 versus 0.146), but external validation on a 10,000-scenario archive reversed this ordering (0.104 versus 0.126; 0.128 for random 45-run subsets), and the GP-selected design outperformed all 20 random subsets. LOOCV on adaptively selected points is therefore an unreliable basis for comparing acquisition rules. SHAP analysis of the archive showed that slow-charging rate and arrival intensity dominate waiting time, throughput, and dissatisfaction, while tariffs mainly affect revenue. GP metamodels fitted to 30 and 45 runs recovered these drivers more closely than random subsets of the same size, so a small, well-spread simulation design can support credible screening of charging-network planning scenarios.