Multi-wind farm short-term wind ramp event warning via temporal representation learning and imbalance-aware optimization

Wind power ramp events threaten power-system operation, making short-term warning important. This study develops and evaluates an application-specific Multi-Wind-Farm Temporal Convolutional Network (MWF-TCN) that uses a shared temporal encoder and classifier across six farms while processing each farm’s input sequence separately. We compare it with non-temporal and graph-based alternatives. With a chronological, leakage-controlled split, the one-hour warning F1-score was 0.4021 ± 0.0055 across three random seeds, compared with 0.3953 for XGBoost. The paired F1 difference was not statistically resolved. For one matched-seed comparison, adding the investigated adaptive graph convolution reduced F1 from 0.3962 to 0.3438. These results support competitive recent-history warning performance and show that this graph design did not help in the evaluated configuration; they do not establish general superiority over XGBoost or spatial modeling methods. False alarms and missed events remain operational concerns.