Improved short-term significant wave height forecasting using ensemble empirical mode decomposition coupled with linear regression

Short-Term significant wave height (SWH) forecasting is essential for many wave energy-related tasks. SWH forecasting techniques may be clustered into three categories: namely, classical, statistical, and data-driven techniques. Data-driven techniques offer less computational burden than classical numerical forecasting methods and improved forecasting capabilities than statistical methods. However, deep learning techniques - even though offering state-of-the-art accuracy in forecasting - long training time and un-interpretability prevent its mainstream adoption. These issues can be mitigated by proper data preparation. this paper proposes a fast and efficient technique that offers excellent forecasting accuracy and fast training time. A combination of Ensemble-Empirical-Mode-Decomposition and Linear Regression (EEMD-LR) is used to forecast the 1 Hr. SWH, with input features’ lag chosen using Bayesian optimization. Several National Buoy Database Center (NDBC) buoys were used to validate the model. The proposed technique outperformed many published techniques, achieving an average improvement of 3.5% and 50.9% in coefficient of determination and mean absolute error metrics, respectively eight state-of-the-art deep learning techniques. Additionally, it offered a short training time with fewer data required for training.