Physics-constrained transformer for wind power forecasting

Wind energy serves as a cornerstone in the global transition toward carbon neutrality, standing among the most vital renewable energy resources. Accurate wind power forecasting is critical for efficient grid integration and market operations, as a 15% error in prediction can result in profit losses of up to 13.8% for energy market participants. Nonetheless, precise wind power forecasting remains a challenging task, primarily due to the presence of high levels of noise in wind speed forecasts, which significantly undermines both the accuracy and robustness of predictive models. To address this issue, this study proposes a novel method named PCT (Physics-Constrained Transformer) for wind power forecasting. PCT integrates the efficient Transformer architecture with domain knowledge of wind power curves, aiming to not only enhance forecast accuracy but also maintain strong robustness in the presence of high-noise conditions. By embedding physical constraints into the Transformer structure, PCT ensures that its outputs conform to the expected probabilistic behavior of real-world wind power generation, thereby improving resilience against noisy input data. Experimental results based on operational data from 25 distinct wind turbines demonstrate the effectiveness of PCT. Under low-noise scenarios, the proposed method achieves an average performance improvement of approximately 15% over the previous state-of-the-art model. More notably, under scenarios involving different levels of input noise, PCT exhibits considerably better stability and accuracy compared to the standard Transformer model trained solely with MSE loss, highlighting the effectiveness of integrating domain knowledge into the Transformer model.