Fine-grained ultra-short-term wind power forecasting based on Temporal Fusion Transformers integrated with turbine power time series clustering

Ultra-short-term wind power forecasting (0–4 h, 15 min resolution) is crucial for turbine operation, power grid stability, and energy market efficiency. However, conventional forecasting approaches often model wind farms as homogeneous systems, overlooking turbine-level heterogeneity arising from geographic dispersion, wake effects, and equipment variability. This simplification obscures localized power fluctuation patterns and constrains forecasting accuracy. To address the limitation, this study proposes KSC-TFT, a fine-grained forecasting framework based on the divide-and-conquer principle, which integrates shape-based turbine clustering (KSC) with parallelized temporal modeling using the Temporal Fusion Transformer (TFT). The KSC module adopts a multi-step process to cluster turbines with similar temporal patterns, capturing phase-invariant characteristics and stabilizing cluster assignments via a consensus matrix. Subsequently, TFT performs parallel forecasting across clusters, leveraging attention-based modeling of long-range dependencies and static covariate encoding. This dual-stage design enhances forecasting accuracy through pattern amplification (reinforcing shared fluctuations within clusters) and error source isolation (reducing cross-turbine interference). The proposed framework supports both deterministic and probabilistic ultra-short-term wind power forecasting, with flexible integration of numerical weather prediction (NWP) data. Case studies demonstrate significant improvements over baseline models, with average reductions of 50.2% in NMAE and 50.8% in NRMSE for deterministic forecasts, and 31.3% and 28.4% reductions in normalized CRPS and pinball loss for probabilistic outputs. These results confirm the framework’s robustness and scalability in capturing spatiotemporal variability.