Physics-informed LLM-guided dynamic spatio-temporal graph deep network for short-term wind farm power forecasting

Wind energy plays a critical role in meeting global electricity demand and decarbonisation targets, accounting for a substantial and growing share of renewable power generation worldwide. However, the inherently intermittent, non-stationary, and spatially coupled nature of wind severely complicates the reliable integration of wind farms into power systems. In particular, short-term wind power forecasting at ultra-short (10-minute) and short (one-hour) horizons is essential for real-time dispatch, reserve allocation, and grid stability, yet remains challenging due to wake interactions, regime shifts in wind direction and speed, sensor noise in supervisory control and data acquisition, and dynamically evolving turbine-to-turbine dependencies. To address these challenges, a Physics-Informed directionality-aware Large Language Guided Dynamic Spatio-Temporal Graph Deep Model is proposed for short-term wind farm power forecasting. The proposed framework integrates physics-informed power-curve cleaning, dynamic graph construction guided by a large language model, and spatio-temporal graph convolutional networks to explicitly model regime-dependent, directed turbine interactions. Unlike conventional static or correlation-based graphs, the large language model infers context-conditioned adjacency structures that adapt to changing wind regimes while preserving interpretability and reproducibility through schema validation and fallback mechanisms. Extensive experiments are conducted on a large-scale, high-resolution dataset from a multi-turbine wind farm, benchmarking the proposed framework against 14 well-known recurrent deep learning, transformer-based, and hybrid ensemble models with two recent LLM-based forecasting baselines. For 10-minute-ahead forecasting, the proposed model (LLXSTG) substantially outperforms the JST-Transformer, achieving 66.5% and 67.9% reductions in mean absolute error and root mean squared error, respectively, together with a 19.3% improvement in R-value. When compared with the advanced ensemble learning baseline, the proposed model yields 32.6% lower mean absolute error, 26.3% lower root mean squared error, and a 4.1% increase in R-value, demonstrating clear gains even against strong non-transformer architectures. At the one-hour-ahead horizon, the proposed model maintains its superiority under increased forecast uncertainty, reducing average absolute error and root mean squared error by 55.2% and 54.2%, respectively, relative to JST-Transformer, while improving the R-value by 11.5%. Compared with the advanced ensemble, the proposed model achieves 35.0% lower average absolute error, 29.6% lower root mean squared error, and a 4.9% increase in correlation. These consistent improvements across both ultra-short and short-term horizons highlight the effectiveness of integrating causality awareness, large language-guided dynamic graph construction, and spatio-temporal learning for robust wind farm power forecasting.