Harnessing Convolutional Neural Networks in Forecasting Significant Wave Heights from MEMS Accelerometers

Machine learning tools are increasingly handy for filling gaps in oceanographic buoy data during downtime by accurately predicting oceanographic parameters within a reasonable lead time from hindsight. A plethora of ML models have been explored for oceanographic applications. As a first step toward applying ML to our operational buoy, we explore a convolutional neural network (CNN) model for forecasting significant wave heights from a MEMS accelerometer mounted on the buoy. First, we determined significant wave heights using the one-third of the mean displacements, 4 standard deviations, and the spectral density technique, applying the Fourier transform to the detrended displacement time series to assess data quality and reliability. Next, we used 10- and 4-day training and test sets, respectively, at a 15-minute data rate. We built a 1-D CNN model that successfully predicted significant wave heights for four consecutive days, with the model evaluation showing RMSE, MAE, and MAP of 0.1 cm, 1.2 cm, and 2.6 cm, respectively, and R2 scores of 96% and an adjusted R2 of 96%.