A Direct Classification Approach for Reliable Wind Ramp Event Forecasting under Severe Class Imbalance

Decision support systems are essential for maintaining grid stability in low-carbon power systems, such as wind power plants, by providing real-time alerts to control room operators regarding potential events, including Wind Power Ramp Events (WPREs). These early warnings enable the timely initiation of more detailed system stability assessments and preventive actions. However, forecasting these events is challenging due to the inherent class imbalance in WPRE datasets, where ramp events are less frequent (typically less than 15% of observed events) compared to normal conditions. Ignoring this characteristic undermines the performance of conventional machine learning models, which often favor the majority class. This paper introduces a novel methodology for WPRE forecasting as a multivariate time series classification task and proposes a data preprocessing strategy that extracts features from recent power observations and masks unavailable ramp information, making it integrable with traditional real-time ramp identification tools. Particularly, the proposed methodology combines majority-class undersampling and ensemble learning to enhance wind ramp event forecasting under class imbalance. Numerical simulations conducted on a real-world dataset demonstrate the superiority of our approach, achieving over 85% accuracy and 88% weighted F1 score, outperforming benchmark classifiers.

💡 Research Summary

The paper tackles the problem of forecasting Wind Power Ramp Events (WPREs) – sudden, large variations in wind farm output that can jeopardize grid stability – by framing it as a multivariate time‑series classification task rather than the more common indirect approach that first predicts power and then derives ramps. Recognizing that WPREs are rare (typically <15 % of observations) and that conventional machine‑learning models tend to be biased toward the majority “normal” class, the authors propose a comprehensive pipeline that explicitly addresses class imbalance and the real‑time nature of ramp detection.

Key contributions include: (1) a data‑preparation stage that extracts statistical features (mean, variance, rate of change, etc.) from the most recent l power observations and masks ramp labels that are not yet finalized, thereby encoding the inherent uncertainty of ongoing ramps without treating them as missing values; (2) a formulation of the forecasting problem where each training instance consists of a flattened 2 × l matrix (power and masked ramp information) and a target label representing the ramp type at horizon h (usually h = 1); (3) the use of majority‑class undersampling combined with the EasyEnsemble ensemble method, which builds multiple sub‑ensembles on different undersampled subsets of the majority class and aggregates their predictions, preserving minority‑class information while still exploiting the bulk of the data; (4) seamless integration with a real‑time ramp‑identification algorithm (Swinging Door Algorithm, SDA) that signals when a ramp trend changes, triggers the masking of the last observation, and initiates a new event window. The model is trained offline on historical records and then applied online without a feedback loop, making it compatible with existing operational tools.

The authors evaluate the approach on a real‑world dataset collected from wind farms in Belgium and Italy, spanning several years of 10‑minute resolution measurements. The dataset exhibits a severe class imbalance with only about 12 % of samples belonging to various ramp categories (up‑ramp, down‑ramp, etc.). Experiments compare the proposed EasyEnsemble‑based classifier against baseline methods such as Random Forest, XGBoost, LSTM, and CNN classifiers. Results show that the proposed method achieves over 85 % overall accuracy and a weighted F1 score of 88 %, substantially outperforming the baselines (the best baseline reaching ~79 % accuracy and ~80 % weighted F1). Importantly, recall for the minority ramp classes improves markedly (e.g., up‑ramp recall rises from ~60 % to >80 %). The masking strategy accounts for only ~7 % of the data, yet it does not degrade performance, confirming that the model can learn from uncertain, partially observed events.

The discussion highlights the practical benefits of directly forecasting ramp events without relying on external meteorological inputs or intermediate power forecasts, thereby reducing error propagation inherent in indirect methods. The EasyEnsemble technique proves effective for handling severe imbalance, while the masking of ongoing ramps provides a realistic representation of the information available to operators in real time. Limitations include the lack of online model adaptation (requiring periodic retraining to cope with concept drift) and the sensitivity of the masking constant to domain‑specific definitions of “ongoing” ramps. Future work is suggested in three directions: (i) incorporating incremental learning to update the model continuously as new data arrive; (ii) refining the uncertainty representation by dynamically adjusting mask values or employing probabilistic forecasts; (iii) extending the framework to multimodal data (e.g., incorporating weather forecasts, storage state‑of‑charge) and validating the approach in a live control‑room pilot.

In conclusion, the study presents a novel, class‑imbalance‑aware, direct classification framework for WPRE forecasting that integrates smoothly with existing real‑time detection tools, delivers superior predictive performance on real wind‑farm data, and offers a solid foundation for further research and operational deployment.