A Hybrid Linear–Gaussian Process Framework with Adaptive Covariance Selection for Spatio-Temporal Wind Speed Forecasting

Accurate wind speed forecasting is critical for renewable energy planning and meteorological studies. However, wind behaviour is complex due to the presence of synoptic systems, terrain effects, and turbulence. This paper proposes a new model that uses a linear regression mean model and a Gaussian process for residual modelling. The monitoring stations were clustered by geographic coordinates and elevation. The Hopkins statistic was employed for cluster validation, while silhouette values were employed for cluster quality validation. It was found that for stations at high elevation located in the interior (Cluster 2), the GP model for residual modelling consistently improved wind forecast accuracy by up to 16.3%. However, for coastal stations at low elevation (Cluster 1), the GP model was not effective for residual modelling. This proves that the accuracy of GP residual modelling depends to a large extent on the wind regime.