Research On Instantaneous Surface Elevation Prediction Algorithm Based On Neural Network

The marine environment is a basic research object in the field of shipbuilding and ocean engineering.Accurate,reliable and fast wave forecasting has vital scientific research value,safety significance and economic significance for marine operations and marine transportation.However,in fact,the nonlinear and non-stationary nature of the wave elevation time series of irregular sea waves makes autoregressive forecasting extremely challenging.Numerical simulation based on Computational Fluid Dynamics(CFD)is one of the effective methods for time-domain analysis of wave evolution,but the stability and accuracy of simulation results are easily affected by various simulation parameter configurations,and there are physical Model error,discrete error and iteration error,etc.More importantly,in order to ensure the accuracy of the solution,the calculation efficiency of simulation software simulation is often not ideal,and a conventional desktop device is often not competent.In recent years,artificial intelligence technology has developed vigorously,and artificial neural network models have strong data fitting capabilities and high real-time reasoning capabilities.At present,artificial intelligence technology has been deeply integrated into the production and research practice of various industries,and has become a new scientific research paradigm after theory,experiment and computational analysis.This also opens up new solutions for analyzing wave evolution and making wave forecasts.This thesis proposes a wave elevation instantaneous forecast model based on neural network and related technologies,and performs autoregressive forecast on the irregular wave elevation data simulated by CFD industrial software,and analyzes different improvement and optimization schemes of the forecast model,aiming at In the development of a wave elevation forecasting method in the field of shipbuilding and ocean engineering that has sufficient generalization and can adapt to small sample size training and forecasting tasks.The instantaneous wave elevation forecasting method based on the autoregressive forecasting strategy proposed in this thesis is also suitable for scenarios such as ship and ocean platform motion forecasting,and it also provides further understanding of potential alternatives to traditional numerical models for simulating water wave evolution.In this thesis,the CFD commercial software STAR-CCM+ fluid analysis module is used to carry out the research and analysis of the above-mentioned water wave prediction based on various machine learning and deep learning algorithms.The specific work content is as follows:(1)First,conduct a review of relevant domestic and foreign literature,analyze the development history and research status of computational fluid dynamics and numerical wave tank,and give an overview of machine learning and its technology applications.(2)Based on the relevant theories of hydrodynamic analysis,the relevant numerical simulation theories involved in this thesis are expounded.Introduce the basic governing equations of fluid motion,wave making and wave dissipation technology,overlapping grid technology,etc.(3)A more comprehensive introduction to machine learning technology,and an analysis of the theoretical basis of machine learning and deep learning algorithms used in this thesis.At the same time,sufficient and effective wave elevation data samples are generated through numerical simulation,the time series characteristics of the collected wave elevation data are analyzed,and the selection of models and the design of time series prediction algorithms are introduced in detail.(4)Preprocess the collected wave elevation sequence data,then split the sequence,make samples,construct sample sets,divide training sets and verification sets,and then use the four sets of neural network prediction algorithms initially designed,which are not sufficient for the sample size training and prediction experiments were carried out on the data set,and the algorithm design with the least dependence on big data and the most generalization ability was selected.Then use the AE initialization method to initialize the selected model,and then carry out small sample size training and prediction experiments,and verify the effect of the initialization method on improving the model’s prediction performance.(5)Integrate the model selected in the above experiment with another machine learning algorithm e Xtreme Gradient Boosting(XGBoost),and conduct small sample size training and prediction experiments on the integrated model to test its generalization Performance improvements.Then let the integrated model fully learn on a sufficient amount of irregular wave data to verify the prediction performance of the model in the data-driven mode.Finally,the model is transferred to the small sample size training and prediction tasks of strongly nonlinear deformed waves to verify the portability and practicability of the model.