Online probability density prediction of wind power considering virtual and real concept drift detection

Renewable energy sources such as wind are inherently variable, making accurate prediction crucial for effective power system operation. However, the inherent volatility and variability in wind power data often result in concept drift, which undermines prediction accuracy and poses significant challenges to power system operation. To address this issue, this paper proposes an online drift detection and adaption combined with quantile regression long short-term memory network (O-DDA-QRLSTM) for probabilistic wind power prediction. This method uses QRLSTM for quantile prediction of wind power and incorporates online learning to detect concept drift in the data stream. Different detection methods are developed for virtual drift and real drift. Kullback-Leibler (KL) divergence of data distribution changes to detect virtual drift and the changes in the Continuous Ranked Probability Score (CRPS) of model predictions to detect real drift. The model parameters are updated to learn new features in the data stream, and Kernel Density Estimation (KDE) is employed to obtain probabilistic density prediction results. Comparison results show that the proposed method effectively handles concept drift and achieves probabilistic density prediction of future wind power, thereby reducing the uncertainty of future information.