Project Description
Real time wave buoy observations inform a range of marine operations, and if buoys are deployed sufficiently upstream, can provide several hours of early warning for an arriving swell or storm. However, determining the exact arrival time and transformation in height and other wave parameters between offshore buoys and nearshore locations can be complex and difficult to predict. A common existing approach to provide a short-range (<12 hr) forecast of wave conditions at coastal locations is to use the offshore buoy observations to initialise a spectral numerical wave model, with new simulations being run each time new buoy observations arrive. Spectral models, while being very valuable tools, are reasonably computationally expensive over regional scales (e.g. the NW Shelf) and either do not contain the necessary physics or have sufficient resolution to accurately predict wave transformation in complex coastal settings (e.g. around and in-between islands). To avoid rerunning the model each time new buoy observations come in, a hybrid approach is to use the buoy observations along with a series of wave model simulations to develop a spectral transfer function to predict waves at a given coastal location given a set of offshore observations (i.e. a ‘look-up table’). However, while more computationally efficient than continually running a numerical model, this approach does not resolve the issues with spectral models in terms of their physics nor does it incorporate other important factors (e.g. tides) that may impact nearshore wave conditions.
The prototype system developed for this project utilised observed wave conditions at Woodside’s network of offshore buoys and predicted (local) tidal levels, in order to produce a ‘now-cast’ prediction of the wave conditions at the LPG Jetty in Mermaid Sound (Figure 1).
A key focus of the project was to train the ML system and make all predictions in the spectral space, using the complete 2-dimensional spectra (energy as a function of frequency and direction). The 2D spectra complicates the system both in terms of the level of complexity in the ML model and the computational resources required (i.e. each observation/prediction is a matrix rather than single values for the wave height, period and direction). However, the advantage of the spectral approach is that it provides a much richer data set to interrogate for operational purposes (e.g. predicting vessel motions) – particularly when multiple swell/sea spectral peaks are evident, as is often the case in Mermaid Sound.
This project has shown that ML can provide realistic real-time wave forecasts (now-casts) when fed with sufficiently upstream spectral wave observations. The spectral ML model appears to resolve the individual frequency and direction peaks, but has an overall tendency to underestimate the energy and thus wave height (e.g. Figure 6).
Additional work could fine tune the model and develop a deployable tool for Woodside. If an operational ML model were to be successfully developed, industry would benefit from more accurate shore-range wave forecasts (that would require less computational resources). These improved forecast would directly benefit a range of marine operations including ship berthing and loading.
Figure 1. (A) Woodside’s network of wave buoys (red circles) on the Northwest Shelf. (B) Inset showing Mermaid Sound (black rectangle in A) with the location of the LPG jetty and associated wave buoy (green star), King Bay tide gauge (blue pentagon), and additional wave buoy (red).