Data-Driven Harmonic-State Identification from Load Composition and Operating Conditions

Total harmonic distortion (THD) measurement is expensive to deploy at scale across the many feeders of a university microgrid, yet aggregate appliance-composition and operating data are routinely logged and cheap to collect. We ask whether such context is sufficient to (A) identify harmonic operating states without observing THD, (B) estimate THD magnitude, (C) drive an adaptive policy that requests a THD measurement only when it is worth its cost, and (D) whether a sequential reinforcement-learning (RL) formulation improves on one-shot selective classification when the acquisition environment is reconstructed retrospectively. Because no field dataset accompanies this study, we generate a fully synthetic 456-observation, 19-feeder, 24-hour dataset whose structure mirrors a real university-microgrid workbook, including injected spreadsheet-formula anomalies and a genuine feeder-mislabeling artifact that must be evidence-gated before use. On this data we run the complete methodology end to end: cluster-aware bootstrap inference, feeder-grouped nested cross-validation across thirteen classifier and fourteen regressor families, a decision-theoretic selective-acquisition study with a mandatory random-acquisition baseline, and fitted-Q-iteration RL from five starting-point axes on a staged-information Markov decision process. We find that no single nonlinear-appliance-penetration ratio has a pooled relationship with THD that survives feeder-cluster bootstrap resampling, while per-feeder relationships range from strongly negative to strongly positive; that feeder-grouped classifiers identify high-THD states on unseen feeders well above chance; that a nested risk-threshold acquisition policy beats random acquisition at every measurement budget; and that, contrary to the naive expectation that a well-audited one-shot policy is hard to beat, retrospectively trained RL agents are competitive with, and at matched rate and a shared base classifier, sometimes exceed, the honest nested supervised policy. We report this as a methodology and reproducibility artifact: every number is computed from the accompanying code and released synthetic data, and all conclusions are scoped to that synthetic setting rather than claimed as field findings.