Harmonic distortion at low-voltage points of common coupling (PCCs) is an operational concern as power-electronics-interfaced loads proliferate across distribution networks. This study investigates whether power-quality index telemetry, rather than continuous waveform measurements, can support probabilistic forecasting of voltage total harmonic distortion (UTHD) and its margin to an IEEE 519-2022 reference boundary. A telemetry-first framework is developed using exported power-quality indices at 10-min intervals. A multi-output histogram-based gradient-boosting regressor forecasts per-phase UTHD at 10-, 30-, 60-, and 120-min horizons and is benchmarked against six baseline methods. Adaptive split-conformal calibration is applied to construct prediction intervals, with finite-sample marginal coverage interpreted under the exchangeability assumption. Using an 11,088-sample data-centre dataset spanning 77 days, the primary phase-B target achieves a mean absolute error of 0.274 percentage points at the 10-min horizon, representing a 20% reduction relative to persistence (p < 10−4). Frozen-model transfer to two independently instrumented external sites is also evaluated, with target-site conformal recalibration using labelled calibration subsets. The results provide empirical evidence that index-only telemetry can support multi-horizon UTHD forecasting with uncertainty intervals, while distributional shift provides a useful diagnostic for cross-site transfer.
Increasing deployment of power-electronic loads has intensified the need for short-horizon forecasting of voltage harmonic distortion at low-voltage points of common coupling. This study develops a telemetry-first framework for forecasting phase-wise voltage total harmonic distortion (UTHD) and its compliance margin using routinely retained 10 min power-quality indices. Independent histogram-based gradient-boosting regressors (HGBRs) were trained for 10, 30, 60, and 120 min horizons using leakage-controlled chronological validation and were benchmarked against persistence, seasonal and climatological references, Ridge, autoregressive integrated moving average (ARIMA), bidirectional long short-term memory (BiLSTM), target-history-only HGBR, XGBoost, and LightGBM. Forecast uncertainty was quantified using symmetric conformal prediction with calibration-only selection of adaptive residual parameters. Across the three Córdoba phases, the common-feature HGBR achieved mean MAEs of 0.198, 0.225, 0.232, and 0.258 percentage points at 10, 30, 60, and 120 min, respectively, and consistently outperformed persistence. The selected conformal configuration achieved mean empirical coverages of 90.55–88.77% for nominal 90% intervals and 95.15–94.17% for nominal 95% intervals across increasing horizons. These findings show that index-level telemetry can support accurate multi-horizon UTHD forecasting with empirically calibrated predictive uncertainty for forecast-informed harmonic monitoring.