NCF: A physics-constrained pre-trained method for short-term wind power forecasting

A physics-constrained pre-trained framework, termed NWP-Chronos-Fusion, is proposed for short-term wind power forecasting. The framework combines a physics-informed fusion module grounded in the theoretical power curve, a Chronos-Bolt-based trend prediction module, an NWP-driven instantaneous power estimation module with bias correction, and a multi-segment nonlinear fusion module. The framework is distinguished by three design principles: integrating a pre-trained time-series foundation model with targeted fine-tuning, incorporating a piecewise-weighted loss function with physical consistency constraints, and establishing a dual-pathway architecture that jointly captures long-term trends and instantaneous meteorological responses. Experimental validation across mountainous, offshore, and flat-terrain wind farms demonstrates that, against state-of-the-art deep learning baselines, mean absolute error is reduced by 29.77%–42.27%, root mean square error by 28.31%–36.18%, and R2 increases by 144.00%–280.77%. These results confirm that NCF achieves improved accuracy and physical interpretability across diverse terrain conditions.