We propose the use of techniques from Machine Learning for the prediction of tidal currents. The classical methodology of harmonic analysis is widely used in the prediction of tidal currents and computer algorithms based on the method have been used for decades for the purpose. The approach determines parameters by minimizing the difference between the raw data and model output using the least squares optimization approach. However, although the approach is considered to be state-of-the-art, it possesses several drawbacks that can lead to significant prediction errors, especially at locations of fast tidal currents and ’noisy’ tidal signal. In general, careful selection of tidal constituents is required in order to achieve good predictions, and the underlying assumption of stationarity in time can restrict the applicability of the method to particular situations. There is a need for principled approaches which can handle uncertainty and accommodate noise in the data. In this work, we use Gaussian process, a Bayesian non-parametric technique, to predict tidal currents. The overall objective is to take advantage of the recent progress in machine learning to construct a robust yet efficient algorithm. The development can specifically benefit the tidal energy community, aiming to harness energy from location of fast tidal currents.
