Spatially varying coefficient (SVC) models allow variable relationships to change over space for clear interpretability, but traditional Bayesian Gaussian Process approaches face severe computational bottlenecks on large datasets. By integrating the Vecchia approximation into a Bayesian SVC framework, this method dramatically reduces computational complexity while accurately mapping spatial heterogeneity—even in challenging real-world scenarios with high data censoring and irregular sampling designs.

By training neural networks on simulated Gaussian random fields, this framework replaces slow traditional optimization with near-instantaneous parameter estimation and uncertainty quantification for large spatio-temporal datasets. Applied to solar radiation forecasting, it effectively manages temporal non-stationarity while delivering accurate predictions in a single forward pass.