A Dynamic Ensemble Deep Randomized Neural Network Using Deep Autoregressive Features for Wave Height Forecasting with Missing Values

Abstract

Wave energy is an essential part of sustainable energy. Precise forecasts of wave height assist in the reliable control of wave energy converters and the intelligent operation of electricity generation. However, the severe and extreme environment poses a significant challenge for accurate sensor recording, resulting in a huge number of missing values at random. The missing values exist in multiple explanatory variables, significantly deteriorating the performance of the classical machine learning models. This article aims to enhance the accuracy of significant wave height forecasting with data imperfections by proposing a flexible dynamic ensemble framework and an ensemble deep randomized neural network. First, the proposed dynamic ensemble framework disaggregates the whole forecasting task into multiple subtasks based on the number of missing values. For each subtask, any missing values imputation method can be employed due to the strong flexibility of the dynamic ensemble framework. This article proposes a novel ensemble deep random vector functional network with deep autoregressive features (DARedRVFL) as a base learner under the dynamic ensemble framework. The deep autoregressive features assist in extracting temporal features. Finally, combining the proposed dynamic ensemble and DARedRVFL achieves the minimum average rankings.