This paper presents a novel spatio-temporal forecasting approach using a Two-Step Linear Regression (Two-Step LR) approach. In the first step, the complete set of features is divided into subgroups and Linear Regression is applied to each subgroup to obtain individual regression values. In the second step, these values serve as inputs for another Linear Regression, producing the final prediction. We demonstrate that Two-Step LR is a natural extension of the Two-Step Linear Discriminant Analysis (Two-Step LDA) originally developed for high-dimensional EEG data classification by leveraging the spatio-temporal structure, particularly the separable covariance matrix. In this study, we demonstrate that wind speed data exhibit an analogous separable structure. This property facilitates the effective implementation of the Two-Step LR method through the adoption of the feature grouping strategy originally developed for Two-Step LDA. Empirical findings indicate that the proposed Two-Step LR model achieves the state-of-the-art performance in wind speed forecasting. This task is crucial for wind energy applications, such as airborne wind energy systems in the field of aerial access, due to the inherent variability and unpredictability of wind dynamics.