Ocean wave characteristics prediction and its load estimation on marine structures: A transfer learning approach

Statistical and stochastic approaches, commonly used to estimate wave induced loads in the marine and coastal structural analysis and design of offshore vessels and platform, require accurate sea state or wave characteristics, which are usually estimated using computationally intensive numerical wave models. While advances in machine learning helps to overcome the computational challenges of these approaches, they still need measurement of historic data at the region of interest to help prospective estimation of wave loads. This calls for deployment of sensors and allied resources for continuous monitoring and collection of streaming data. Transfer learning, on the other hand, provides a framework to derive representations from an existing region to estimate the characteristics in the region of interest. Deep Belief Networks are excellent candidates for latent representation of the relationship between the different sea state characteristics and the induced loads. In this paper, the authors first train a DBN based wave characteristic prediction model to predict the sea state characteristics (namely, significant wave height Hs, dominant or peak wave period Tp and average or zero crossing period Tz) using data from 12 stations at 3 geologically different regions (the Gulf of Mexico, the Korean Region and the UK region), using data between the period Jan 1, 2011 and Dec 31, 2014. The authors validate the model using data in these regions between Jan 1, 2015 and Aug 30, 2015. They then compute the effect of the wave characteristics, namely, the wave-drift force and moment on 12 different marine structures with both regular and irregular waves. Finally, the authors transfer the DBN representations to predict the sea state characteristics in Irish region for the same period. Evaluations against ground truth and comparisons with state-of-the-art methods show the superior transfer learning performance of the DBN and accurate computations of the wave drift force and moments.

Language

  • English

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01681264
  • Record Type: Publication
  • Files: TRIS
  • Created Date: Aug 6 2018 3:09PM