C-MixNet: Regime-aware clustering and MDN-augmented LUBE for uncertainty-aware wind power forecasting

Wind power forecasting under uncertainty remains a critical challenge due to non-stationary patterns, regime shifts, and fluctuating dynamics across time. Existing interval-based methods often struggle to generalize across heterogeneous operating conditions, resulting in inconsistent coverage and suboptimal reliability. This paper presents C-MixNet, a hybrid framework that integrates regime-aware clustering, Mixture Density Network (MDN), and Lower Upper Bound Estimation (LUBE) for robust Prediction Interval (PI) generation. Using three years of SCADA data from a Vestas V52 turbine in Ireland, we first segment the dataset into regimes using an ensemble of Gaussian Mixture Model (GMM), Self-Organizing Map (SOM), and K-Means Clustering Algorithm (K-Means), resolving label disagreement via majority voting and measuring agreement with Cohen’s kappa scores. For each cluster, we train dedicated MDN models to extract probabilistic forecasts and embed their outputs, including component weights, means, and variances, into the feature space, enabling LUBE to learn enriched uncertainty-aware representations. To manage the dimensionality introduced by MDN features, SHapley Additive exPlanations (SHAP) based feature selection is employed, although reduced models underperform. We then incorporate MDN-derived predictive bounds as additional features, which substantially improve LUBE performance. Conformal prediction is further explored for calibrated interval estimation but shows limited improvement under the considered evaluation criteria. Additional validation examines temporal dependency, cross-turbine applicability, and comparison with a conventional probabilistic benchmark. Ultimately, the MDN-augmented LUBE framework achieves the best trade-off in Prediction Interval Coverage Probability (PICP), Prediction Interval Normalized Average Width (PINAW), and Coverage Width-based Criterion (CWC) across clusters. Extensive experiments with ten repeated trials validate the consistency and robustness of the proposed approach. This work contributes a modular, regime-adaptive, and interpretable framework for uncertainty-aware wind power forecasting, with strong potential for real-world deployment in energy markets and operational dispatch systems.