Replication package for "What Aggregate Metrics Hide in Electric Vehicle Charging Control: A Distributional Evaluation on a Two-Network Digital Twin"

This is the replication package for a study of what aggregate reporting conceals in electric vehicle (EV) charging control. Evaluations of EV charging control are dominated by aggregate metrics: total energy delivered, aggregate demand completion, mean curtailment and overload counts. Drivers, however, experience individual sessions, and a controller operating under a binding capacity constraint must decide whose charge to reduce. This package contains the per-session evidence behind a distributional evaluation of six controllers on a predictive digital twin driven by real charging demand from two independent networks. CONTENTS Data. Per-session logs covering 129,864 rows, being 21,644 held-out sessions evaluated under each of six controllers (an uncontrolled baseline, PID, a forecast-driven proportional throttle, and PPO, SAC and TD3 reinforcement learning agents). Session identifiers are stable across controllers, so all controller comparisons are paired. Hourly logs, a per-controller aggregate summary, and the 1,702-hour forecast evaluation window are also included. Code. Five standalone Python scripts reproducing every number in the paper: aggregate and distributional metrics with 10,000-resample bootstrap confidence intervals, Kendall rank agreement between aggregate and distributional measures, Wilcoxon signed-rank tests on paired service fractions, Lorenz curves and Gini coefficients, request-size quartile analysis, and a variance decomposition of per-session service with cluster-bootstrap confidence intervals. The scripts require only numpy, pandas and matplotlib. Notebook. The end-to-end pipeline that produced the reported run: data loading, quantile LSTM forecaster training, split-conformal calibration, and all six controllers. Figures. The four figures from the paper at 300 dpi. Literature audit. A coding sheet of 20 EV charging control papers with a codebook, recording for each paper whether it reports an aggregate metric, a per-session distribution, an explicit fairness index, a worst-case or tail outcome, and whether its formulation guarantees each vehicle's energy as a hard constraint. Nine included papers all report aggregate metrics; five report per-session distributions; three report a fairness index, all three from the scheduling-theory literature rather than from machine-learning or model-predictive control work. KNOWN ISSUES KNOWN_ISSUES.txt documents the deposit's limitations openly, verified against the included notebook. These are: a curtailment counter that double-counts across waiting hours and is therefore excluded from every reported metric, though it does enter the reinforcement-learning reward; PID gains selected at the lower boundary of the tuning grid; single-seed training for the three reinforcement-learning agents; a deliberately rescaled capacity envelope; proportional within-hour allocation, so all measured inequity accrues across hours rather than within them; the absence of serialised agent weights; and a single-coder, AI-assisted literature audit with eight of twenty candidate papers unscreened. SOURCE DATA The raw third-party inputs are not redistributed. The included files are derived simulation outputs. Rebuilding from source requires obtaining the UK charging transaction datasets and ACN-Data from their original providers under those providers' terms.