Uncertainty Modeling in Significant Wave Height Forecast

Abstract

This thesis proposes different variants of LSTM models for point and probabilisticforecasting of significant wave height (SWH), a crucial component of wave energy.SWH forecasting is challenging due to ocean waves� complex and chaotic nature.The thesis applies different decomposition methods, such as wavelet decomposition(WD), empirical mode decomposition (EMD), and variational mode decomposition(VMD), to enhance the performance of LSTM models. The thesis alsouses a convolutional neural network (CNN) and a genetic algorithm to improvethe feature extraction and hyperparameter tuning of LSTM models. Moreover,the thesis develops a probabilistic forecasting model for SWH using the pinballloss function, which captures the uncertainty and provides confidence intervalsfor the forecasts. The thesis evaluates the proposed models on seven real-worldSWH datasets collected from four different ocean buoys. The results show that theCNN-LSTM model outperforms other LSTM variants in deterministic forecasting,while the probabilistic forecasting model provides reliable and sharp confidenceintervals for SWH.