Multistep Joint Probabilistic Forecasting of Offshore Wind Power: A Confidence-Triggered Clustering Missing-Data Tolerant Model

Accurate and reliable power generation forecasts of offshore wind farm clusters are crucial for the low-carbon operation of multienergy power systems. In practice, measurement data may not always be complete due to various failure issues in data acquisition systems or communication interruptions in harsh marine environments, and missing essential data may significantly reduce the credible prediction accuracy of probabilistic models. To address this problem, this article proposes a novel missing-data tolerant model based on confidence-triggered fuzzy clustering quantile-enhanced transformer (CFCQET). First, a quantile-enhanced transformer-based multistep wind power probabilistic forecasting method is developed, where the predicted values are iteratively updated by conditional confidence expectations. Then, based on the spatio-temporal characteristics of wind farms, a FCM clustering model for offshore wind farms is constructed to divide wind farms with similar power curve attributes for joint modeling. Next, a confidence-triggered strategy is designed for probabilistic power forecasting with missing data under wind farm clusters, where the output interpolated predicted values are used to fill in unobserved input data. Finally, probabilistic prediction tests for twelve offshore wind farms at a time resolution of half an hour. The test results demonstrate that the CFCQET achieves a lower negative form of the continuous ranking probability score (CRPS*), as well as superior sharpness and comparable reliability of the prediction intervals with respect to the benchmarks.