One-day wave forecasts using buoy data and artificial neural networks

The forecasting of wave heights is an essential prerequisite for planning, operation, and maintenance works associated with offshore engineering, navigation, and other activities. The requisite wave height information is currently derived from numerical models which use predicted wind information based on wind-wave relationships. These models sometimes cannot be applied easily at remote locations or on small geographic scales where forcing functions (wind-fields and open ocean boundary conditions) may be unavailable. This paper attempts to forecast waves based on observed wave data only, using the data-driven artificial neural network (ANN) approach. Wave heights with varying lead times of 6 to 24 hours are predicted at five different NDBC buoy locations around the US, viz. two near Alaska (in Prince William Sound), two off the northeast coast (in the Gulf of Maine) and one in the northern Gulf in Mexico (south of Galveston, Texas). Three-layered feed-forward back-propagation networks were used along with the conjugate gradient algorithm. The results show that ANN models perform extremely well at all five locations for the 6 to 12 hour predictions and moderately well for 18 to 24 hours predictions. Online models were also developed using these trained networks and run daily for a period of two months (April-May 2005) to forecast the next day´s wave heights. The results of these models also follow the same level of accuracy achieved in testing these networks as mentioned earlier. When large grid-scale modeling is not possible, buoys that measure wave heights provide the requisite wave information. The present work can be viewed as an attempt to enhance the value of buoy data by providing a forecast. This modeling approach will provide a useful tool for forecasting or supplement wave data especially for locations where buoys have been recently installed or where established wave models are difficult to use due to area limitations.