PV-Aware Multi-Horizon Probabilistic Net-Load Forecasting with Horizon-Delayed Online Conformal Calibration
Mogana Kumaran Sivaraman
*
GAP Inc., San Francisco, CA, USA.
*Author to whom correspondence should be addressed.
Abstract
Rooftop photovoltaics cause household net load to change sign and vary seasonally. We evaluate a PV-aware temporal convolutional quantile network with horizon-delayed conformal calibration on 300 Australian homes and a 299-home portfolio aggregate. The network predicts seven non-crossing quantiles at 48 nominal half-hour forecast horizons. On the original eight selected horizons, PV-TCQN has a 6.82% lower aggregate seven-quantile score (QS7) than the original LightGBM-Q configuration, with the largest gains at short leads; the selection chronology is unknown. Across all 48 equally weighted horizons, PV-TCQN is 0.50% worse numerically, with no statistically conclusive difference. The household comparison, based on one neural seed, is also inconclusive and establishes neither superiority nor equivalence. Rolling-window adaptive conformal inference (RW-ACI) improves the evaluated household home–month–horizon coverage-tail metric relative to rolling CQR, while increasing mean prediction-set length by 4.2–4.4%. Approximately 90% pooled coverage nevertheless coexists with household coverage of 78.64% during high-PV cases and 71.98% during large ramps. These overlapping, outcome-defined retrospective regimes are diagnostics, not conditional-coverage guarantees. A combined in-memory 300-home forecasting call takes 0.520 s median and 0.624 s p95 on one Apple M1 Pro CPU thread. Its 48-row calibration state repeats observed eight-horizon score distributions for implementation scaling; it is not a validated all-48 calibration-quality study or deployment test. The findings support a model- and objective-specific empirical trade-off, while data quality, initialization, and retrospective selection limit generalisation.
Keywords: Net load forecasting, probabilistic forecasting, behind-the-meter photovoltaics, quantile regression, conformal prediction, adaptive conformal inference, temporal convolutional network, smart meter data