Phase-aware photovoltaic forecasting with multi-basis quantum covariance graphs

Abstract

Short-horizon photovoltaic forecasting without future weather information requires effective use of periodic, phase-sensitive, and rapidly changing patterns in recent power histories. This study proposes PhaseQCov-GNet, a hybrid quantum–classical model that converts four phase-sensitive spectral tokens into a sample-dependent signed graph using a four-qubit, two-block data-re-uploading circuit. Connected Pauli correlations measured in the Z/ZZ and X/XX bases define the graph, which is processed by a signed relational mixer and a seasonal-residual decoder. The model uses 48 h of historical power and deterministic calendar variables to jointly predict 1-, 3-, and 6 h-ahead power on the three-zone GEFCom2014 Solar dataset under a leakage-controlled cross-zone protocol. Across the three held-out-zone folds, PhaseQCov-GNet achieved mean RMSE values of 0.092 28, 0.119 41, and 0.131 10 at 1, 3, and 6 h, respectively. It obtained the lowest mean RMSE among the evaluated models at 1 and 6 h, while TinyPatchTST was lower at 3 h. Relative to ClassicalCovGraph, RMSE decreased by 7.59%, 6.73%, and 0.69%; relative to the exactly parameter-matched ClassicalFeatureMapGraph, the reductions were 11.29%, 3.41%, and 1.31%. Graph replacement and whole-set derangement confirmed functional dependence on the generated relations, while ablations showed sensitivity to the controlled- Z layer, multi-basis connected covariance, observable context, seasonal anchoring, and retained spectral phase information. Finite-shot tests also improved with higher sampling budgets. These results support a modest quantum-derived relational inductive bias within the evaluated three-zone, no-numerical weather prediction setting, without establishing computational quantum advantage, causal benefit from entanglement, or universal superiority over classical forecasting methods.