Session: 06-05-01 Marine Hydrodynamics I

Paper Number: 78261

78261 - Machine Learning Based Prediction of Wave-Induced Vessel Response 

Offshore operations are often limited by the vessel response to assure safe and efficient execution of planned work. Wave-induced vessel responses are either inevitable, or often not feasible to mitigate without compromising on economic and environmental concerns, such as fuel consumption and payload capacity. Therefore, accurately describing the wave-vessel interaction is essential for sustainable operations.

Commonly, wave-induced vessel responses are calculated through wave spectra and suitable transfer functions, i.e., response amplitude operators (RAO). This approach is used both for preliminary engineering analysis to establish limiting sea states, and for short-to-medium term prediction of vessel responses based on forecasted sea states during ongoing operations. The latter is crucial for onboard decision support systems and is the focus of the present work.

The vessel response spectrum forms the basis for statistical inference about limiting conditions, since it can be used to derive useful statistics such as standard deviation and zero-crossing period etc. Therefore, accurate response spectrum estimates increase the confidence in the set limiting conditions. Ideally, the response spectrum for a given degree-of-freedom is the product of the corresponding RAO and wave spectrum. However, the accuracy of the estimated response spectra depends on the accuracy of the RAO and the wave spectra, and in real-world scenarios neither of these can be known with certainty. For instance, available RAOs are often linearized with respect to some “reference” sea state and may be outdated as they are developed in the design phase, and they depend on vessel properties that tend to change over time. Furthermore, forecasted wave spectra may suffer from systematic bias and errors, and are inherently uncertain (like any other type of weather forecast).

In the present work, a machine learning model where the “effective” RAO amplitudes are estimated directly from data is presented. That is, measured vessel response spectra and forecasted 2D wave spectra are used to estimate the RAO amplitudes as a 2D surface in the frequency-direction domain, while also correcting for long-term biases that may be present in the 2D wave spectra. It is worth mentioning that measured vessel response spectra are one-dimensional and cannot be decomposed into directional components. Therefore, the naïve approach of dividing the response spectrum on wave spectrum to obtain the RAO is not applicable.

The proposed model is applied on a pipe-laying vessel operating in the Campos/Santos Basin on the coast of Brazil. The vessel was fitted with sensors measuring its heave, pitch, roll, heading, position, and speed continuously. In addition, forecasted 2D wave spectra (ECMWF HRES-WAM) at the vessel location was always provided.  The measurement campaign started August 2020 and is still ongoing as of today (September 2021), providing thousands of data points. (Here, a data point refers to a suitable timespan, e.g., 3 hours of measurement data). The model was trained on the acquired real-world data and used to predict heave and pitch standard deviations for ongoing operations.

The machine learning model proved to be robust and well-suited for a real-world application. It outperformed the predictions based on the conventional RAOs provided by the vessel owner, yielding slightly more accurate forecasts on expected operational limits. Besides providing predictions directly, the model may also be used to evaluate the conventional RAOs, since the estimated RAO (amplitudes) can be visualized side-by-side with other RAOs.

Finally, the proposed approach differs from “common” indirect RAO tuning where an optimal RAO is calculated using numerical methods by altering relevant parameters. In contrast, the present model generates an empirical RAO almost without requiring any knowledge about the vessel hydromechanics. In addition, clever regularizations are introduced to overcome practical issues such as overfitting and divergence. In general, the proposed model will by design performs at least as good as the best available conventional RAO, and therefore, often providing a more accurate description of the wave-vessel interaction.

Presenting Author: Ali Cetin 4subsea

Authors:

Ali Cetin 4subsea
Vegard R. Solum 4Subsea
Cristina Evans Subsea7