Changing the Unpredictable Nature of Internal Tides Through Deep Learning

Abstract

AbstractNonstationary internal tides (ITs) are formed from their interactions with background currents. Harmonic analysis (HA), which cannot be used to estimate the incoherent component, is almost exclusively used method to predict ITs from observations. This remains ITs prediction challenge. In this study, we establish a deep learning framework to predict semidiurnal ITs. The model is established and trained with observed semidiurnal internal tidal currents that are 172 days long, and then ITs over the next 42 days are forecasted. The prediction accuracy is greatly improved using the deep learning framework. The magnitudes of errors using the deep learning framework are approximately 35% of those obtained using HA. Most temporal and spatial variations in baroclinic currents can successfully be forecasted using deep learning. In addition, the kinetic energy and incoherent components of ITs can be accurately predicted. Moreover, the relatively high adoptability of the established deep learning model is shown.