Variable inlet guide vane (VIGV) control is among the most energy-efficient flow regulation methods for axial pumps because it adjusts the inlet swirl angle directly while preserving high hydraulic efficiency. However, unlike rotational-speed control, whose performance curves scale straightforwardly through the affinity laws, VIGV control alters the intrinsic shape of the head-flow (H-Q) curve at each vane angle and therefore requires angle-specific prediction. This study proposes a transition-informed Gaussian process regression (TI-GPR) model that augments standard GPR by explicitly incorporating the gradient sign-reversal point of the S-shaped characteristic curve through adaptive region splitting and sigmoid-based blending. A four-factor evaluation covering CV strategy, training-data sparsity, input dimensionality, and model type shows that, even when only Q–H data are available (2D input), TI-GPR lowers the relative MAPE by 17.525% under sparse interpolation and by 35.938% under extrapolation relative to the baseline GPR model. Adding valve-position information (3D input) improves the accuracy of both models further, and TI-GPR retains its advantage. These results demonstrate that a minimal structural modification embedding the stability-gradient transition boundary can yield substantial predictive gains, particularly in data-scarce regimes.