Transformer-Based Probabilistic Wind Power Prediction and Power Ramping Verification with Error Tolerance

Abstract. Accurate probabilistic wind power forecasts and reliable detection of ramping events are important for the stable operation of wind-integrated power systems. This study implements a Self-Attentive Ensemble Transformer for postprocessing operational wind power ensembles in the Belgian Offshore Zone. This postprocessing delivers corrected ensemble members as output rather than a predictive distribution. The Transformer model achieves near-zero bias in the ensemble mean and a lower Continuous Ranked Probability Score across all lead times than power-curve derived raw ensembles and shows some improvements to Member-by-Member benchmark methods. This model also shows the ability to correct for power overestimation associated with the wake effect. For power ramping, this paper focuses on the predictability of 15 % and 30 % hourly ramping events. The Transformer produces event frequencies that closely align with the observations over all ramping thresholds. Considering that frequent temporal shifts and intensity underestimations limit a fair evaluation of ramping prediction, we introduce an error-tolerant probabilistic verification framework with buffer concepts and score the model by Buffer Brier Skill Score (BBSS). Incorporating a temporal buffer and magnitude tolerance substantially mitigates penalties for minor prediction errors. This reveals that the models are capable of detecting ramping signals, even though they lack strict precision. While uncertainty is inherent in ramping forecasts, the Transformer demonstrates better-calibrated probabilities than the raw ensembles. Overall, the Transformer method improves probabilistic power prediction and provides an informative probabilistic representation of ramping events.