A Mimicking-Avoiding Short-Term Probabilistic Power Forecasting Method for Wave Energies

Abstract

Wave energy is essential for sustainable marine development. However, complex marine weather conditions cause fluctuating wave power outputs, resulting in a mimicking phenomenon in predictions. Moreover, the lack of accurate numerical weather prediction (NWP) data sources aggravates the prediction inaccuracy. To address these obstacles, a series characteristic perception (SCP) method coordinated with an advanced hybrid model, the quantile liberty loss gated recurrent unit kernel density estimation (QLB-GRU-KDE), is proposed for the floating point absorber wave energy system. In the first stage, the SCP method gains prior knowledge via ensemble approaches. In the second stage, a liberty loss function is used to mitigate the mimicking phenomenon. Furthermore, a hybrid model that integrates a GRU and KDE is adopted for the probabilistic forecasting of wave energy generation. In addition, a metric is proposed to evaluate the severity of the mimicking. A case study based on a real-world wave dataset is conducted, where both deterministic and probabilistic prediction approaches are examined. Comparisons with the cutting-edge counterparts reveal that the designed liberty loss effectively mitigates the mimicking issue. The comprehensive performance of the proposed model, including the accuracy, stability, reliability and sharpness in wave power prediction, is validated by multiple metrics.