Decision-aware threshold optimization for photovoltaic ramp event prediction under asymmetric operational costs

Accurate prediction of photovoltaic (PV) ramp events is essential for maintaining grid stability and ensuring reliable operation of renewable-rich power systems. However, conventional evaluation metrics often fail to reflect the operational consequences of forecasting errors under asymmetric cost conditions. This study proposes a decision-aware framework for PV ramp event prediction that explicitly links predictive performance to operational decision quality. A rigorous temporal evaluation methodology combining leave-one-month-out validation with an independent fixed test set is adopted to ensure realistic generalization assessment. Three machine learning models, namely LogitBoost, Random Forest, and Support Vector Machines, are evaluated, with LogitBoost achieving the best overall predictive performance (area under the ROC curve (AUC) = 0.9468, F1-score = 0.7140, and Precision-Recall AUC = 0.7185). The results demonstrate a substantial discrepancy between the F1-optimal threshold and the operational cost-optimal threshold, indicating that conventional metric optimization may lead to suboptimal operational decisions. Post-hoc threshold optimization is further compared with cost-sensitive learning approaches under asymmetric penalties assigned to false positives (100 €) and false negatives (500 €). Although the investigated cost-sensitive learning approaches improve recall, they generate more false alarms and do not achieve the lowest operational cost under the considered evaluation setting, whereas the proposed post-hoc framework yields lower operational expenditure. Furthermore, the analysis reveals a strong dependence of decision quality on weather variability. To address this issue, regime-adaptive decision thresholds are calibrated on an internal forward-chaining validation subset of the training data and subsequently applied unchanged to the independent December test set, thereby ensuring a leakage-free evaluation. On the independent test set, the proposed adaptive strategy reduces operational cost by 3.0% (from 50,400 € to 48,900 €) while improving recall from 0.911 to 0.926 with only a marginal reduction in precision. Within the evaluated PV ramp forecasting setting, the proposed decision-level optimization framework achieves lower operational cost than the investigated cost-sensitive learning approaches, while the low-overhead regime-adaptive thresholding strategy provides additional operational improvements without modifying the predictive model.