Joint probability aggregation for regional wind power forecasting via R-vine copula and Kolmogorov-Arnold networks

• Establishes joint probability aggregation for regional wind power via FCM-R-Vine-KAN. • Integrates KAN with learnable B-splines for adaptive distribution modeling. • Couples R-Vine Copula with FCM spatial decomposition for dependence modeling. • Achieves 16.7% Energy Score improvement via synergistic FCM and R-Vine integration. Probabilistic forecasting for regional wind farm clusters faces significant challenges in capturing complex spatiotemporal dependencies and quantifying high-dimensional joint uncertainty. To address these issues, this paper proposes a novel hybrid framework that synergistically integrates Fuzzy C-means (FCM), Kolmogorov-Arnold Networks (KAN), and R-Vine Copula. Initially, FCM clustering is employed to identify soft spatial patterns, effectively addressing the limitations of traditional rigid clustering in capturing transitional spatial characteristics. Subsequently, KAN is introduced for conditional quantile regression, offering enhanced non-linear approximation capabilities compared to standard Multilayer Perceptron (MLP) to model adaptive marginal distributions. Furthermore, the framework utilizes R-Vine Copula to capture flexible hierarchical dependence structures, thereby overcoming the constraints of conventional multivariate copulas in handling asymmetric tail dependencies. By employing a joint probabilistic aggregation strategy, this approach significantly improves regional uncertainty quantification over direct aggregation. Evaluated on 21 wind farms in Southeastern Australia, the proposed framework demonstrates robust performance: the QR-KAN model reduces the Continuous Ranked Probability Score (CRPS) from 0.0864 (QR-LSTM) to 0.0788 (an 8.80% reduction) for single-site predictions, while the joint aggregation strategy with R-Vine Copula improves the Energy Score from 0.0852 (Direct Aggregation) to 0.0741 (a 13.02% improvement) for regional forecasts. The framework’s ability to faithfully capture spatiotemporal uncertainties directly supports power system scheduling and risk management.