B2MTGP: A Double Bayesian Multi-Task Gaussian Process Method for Wind Turbine Power Forecasting Under Data Uncertainty

Wind power forecasting plays an increasingly important role on the safety, reliability, and stability of the power grid. Due to the fluctuating and intermittent nature of wind resources, it is very challenging to accurately forecast wind power under varying scenarios. This paper proposes a novel Double Bayesian Multi-Task Gaussian Process (B2MTGP) model tailored for wind turbine power forecasting with scarce data. The model leverages both wind speed and limited power data, employing a two-stage Bayesian inference framework: the first stage integrates discrete wavelet packet decomposition to denoise raw signals, enhancing prediction accuracy; the second stage optimizes the multi-task Gaussian process model structure and parameters to account for data uncertainty. A generalized procedure is developed to facilitate automatic implementation of the proposed method for small-sample forecasting scenarios. Comparative experiments using real-world wind turbine datasets demonstrate that B2MTGP significantly outperforms traditional Gaussian process regression and multi-task models, achieving up to 91% improvement in point prediction accuracy. These results highlight the potential of the proposed approach for robust and reliable wind power forecasting under data uncertainty.