Uncertainty-Aware Wind Power Forecasting Using a Cost-Sensitive and Feature Representation-Enhanced LUBE Framework

Precise and dependable wind power prediction is crucial for maintaining grid stability and ensuring operational efficiency. Despite this, current prediction interval (PI) techniques, particularly conventional Lower Upper Bound Estimation (LUBE), focus primarily on optimising statistical width and coverage. These methods typically produce symmetric intervals and fail to account for practical operational risks like storage overflow and backup power shortages. To address this limitation, this study introduces a Cost-Sensitive and Feature Representation-Enhanced LUBE framework. Specifically, the method utilises a newly developed cost function featuring asymmetric penalties tailored to oversupply and shortage risks, which actively directs how the intervals are positioned instead of just assessing them post-hoc. In parallel, PCA, KPCA, t-SNE, UMAP, and Autoencoders are investigated as dimensionality-reduction and representation techniques to construct compact inputs and reduce redundancy. Additionally, a Gated Recurrent Unit (GRU) network forecasts short-term extreme values (MaxPower and MinPower), which are incorporated as auxiliary inputs under multiple historical lag configurations ($t-1$, $t-3$, $t-5$). Through this integrated architecture, the system can simultaneously learn the uncertainty boundaries, the importance of specific features, and its sensitivity to operational risks. Experiments using SCADA data from a Vestas V52 wind turbine demonstrate improved prediction interval quality together with fewer proxy directional violations. At the $t-3$ lag, the Prediction Interval Coverage Probability (PICP) increases from 94.56% to 95.40%, while the Coverage Width-based Criterion (CWC) decreases from 15.74 to 11.99. The proposed framework, therefore, provides a lightweight approach for improving statistical interval quality while incorporating asymmetric directional preferences into LUBE training.