Numerical weather prediction enhanced wind power forecasting: Rank ensemble and probabilistic fluctuation awareness

Numerical Weather Prediction (NWP) is the key to precise wind power forecasting (WPF), which can be enhanced by the NWP correction and scenario partition techniques. However, on the one hand, existing NWP correction techniques may enlarge the volatility of ensemble NWP which disturbs the subsequent WPF. On the other hand, existing scenario partition techniques cannot precisely predict wind power in fluctuating scenarios by assuming NWP is totally reliable. Therefore, this paper proposes a novel NWP enhanced WPF method based on rank ensemble and probabilistic fluctuation awareness. Firstly, Rank Bayesian Ensemble (RBE) method is intended based on the stationary NWP rank, which generates a stable and accurate ensemble NWP. Secondly, a fluctuation scenarios partition framework is devised to establish a fluctuation awareness model with NWP's credibility quantified. The framework works in a three-step manner, including characterization, matching, and inference of wind fluctuation events: respectively as Fluctuation identification and feature embedding (FIGE), Fluctuating mapping algorithm (FMA), and Probabilistic fluctuation warning (PFW). Finally, we incorporate the two enhancement techniques in a forecasting method in the ultra-short-term. A real-world wind farm with four NWP sources data validates the superiority and robustness of the proposed WPF method. The result shows that our method can reduce the four hour-ahead rooted mean square error (RMSE) by 2.16%-4.36% compared to baseline models. Meanwhile, the stability of ensemble NWP and the effectiveness of fluctuation scenario partition are also discussed.

Automated summary

This paper proposes a novel method for enhancing wind power forecasting using numerical weather prediction (NWP) based on rank ensemble and probabilistic fluctuation awareness. Experimental results demonstrate the superiority and robustness of the proposed method in reducing prediction errors compared to baseline models.