Prediction of ocean currents with artificial neural networks

Abstract

The prediction of ocean currents on real time basis is a complex exercise due to the variability of coastal features like topography and bathymetry as well as uncertainty of driving forces and also interaction among various met-ocean parameters. Although numerical methods are commonly used for this purpose, they need large and detailed exogenous information along with high computational resources and at times can have less tolerance to noise and gaps in data. At this backdrop and when site-specific information is sought for, data-driven techniques like artificial neural network (ANN) might appear attractive. Although there are some past applications of ANN to online current forecasting, lower prediction accuracy at higher values and over longer prediction intervals together with unequal accuracy levels for zonal and meridional current components have remained as problems. This paper attempts to address these issues. At two locations in North Atlantic and North Pacific Oceans the ANN-based time series models have been developed to predict currents over time horizons of 1hr to 24 h. After considerable experimentation, it was found that if the input is pre-processed with the help of a carefully selected smoothing technique and if its sequence length is methodically selected, then together with an empirical correction the long interval and extreme predictions significantly improve in case of both meridional and zonal current components.