In recent years, the integration of renewable energy sources in power systems has been increasing. Their inherent unpredictability and output fluctuations pose challenges to secure power system operations and energy market pricing stability. Therefore, an accurate forecast of renewable energy generation is crucial. Several effective forecasting methods that have been applied are based on Machine Learning (ML). A key factor in the application of ML methods is the choice of input features, a task that has become more complex in regional wind power forecasting, where regions can cover entire countries. The proposed method aims to improve forecasting performance by streamlining input features through a data-driven model-agnostic preprocessing technique. This involves splitting the multidimensional numerical weather predictions into subareas and eliminating non-informative subareas. The selection of optimal split and remove parameters is guided by a Bayesian sequential optimisation process, which builds on prior knowledge from previous iterations. The proposed method has been implemented on actual wind power measurements aggregated at regional level for three countries located in Southeastern Europe to demonstrate the effectiveness in improving the performance of popular data-driven forecasting methods. ; ©2024 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. ; fi=vertaisarvioitu|en=peerReviewed|