Significant Wave Height Prediction Based On Extreme Learning Machines And Beyonds

Sea wave is water fluctuation of the ocean surface,and it is a complex and random process.It is important to predict significant wave height accurately and release an extreme wave height warning in advance.These are important to secure the oceanic transportation,oil prospection,oceanic fishery and military activities.Operational significant wave height prediction is based on the wave numerical models,which are a set of ocean dynamic differential equations.The significant wave height prediction based on wave numerical models hardly satisfies the requirements of real applications,because of the model defects,computation relaxation,and error accumulations.On the other hand,machine learning based numerical model correcting methods predict significant wave height from the numerical wave height predictions straightforward.They do not take change rules of the prediction residuals between numerical wave height predictions and the real observations into consideration.In order to predict the significant wave height accurately,we proposed three prediction methods for significant wave height prediction and numerical model correction methods.Firstly,we propose an extreme learning machine based significant wave height prediction method.We commence by forming significant wave height observations to a training dataset.We then train an extreme learning machine using the formed training dataset.Finally,we predict the significant wave height based on the trained extreme learning machine model and multiple continuous historical significant wave height observations.Extreme learning machine has powerful nonlinear mapping capability,and it is efficient because its parameters are not tuned by back propagation and gradient decent algorithms.Experiment validates the effectiveness of our proposed method.Secondly,we develop a single factor wave numerical model correcting method based on residual extreme learning machine.The proposed residual learning method leans the evolution of the prediction residuals between numerical significant wave height predictions and significant wave height observations.We commence by forming the significant wave height prediction residuals to training dataset.We then train the residual learning machine using the formed training data.The trained model is used to predict the future correcting residuals,and the correcting residuals are added to the corresponding numerical significant wave height predictions to obtain high-accurate significant wave height prediction.Our proposed method investigates the hidden parameters which influence the significant wave height prediction,and infer the evolutional principle of significant wave height prediction residuals.Finally,we propose a multiple factor wave numerical model correcting method based on tensor extreme learning machine.The proposed method investigates the effects of air pressure,wind speed,and wind direction on significant wave height prediction residuals.The input of conventional extreme learning machine is a ‘single factor,multiple time' or ‘multiple factor,single time' matrix.Our proposed tensor extreme learning machine expands the input to a‘multiple factor,multiple time' tensor.Experiments validate the effectiveness of our proposed method.Overall,in this thesis,we propose three wave prediction and wave numerical model correcting methods to improve the significant wave height prediction accuracy.Experiments validate the effectiveness of our proposed methods.This project provides new methods for significant wave height prediction,and expands machine learning to a new application scenario.