A combined Gaussian process regression and one-dimensional least squares harmonic method for tidal current prediction

Abstract

Predictions of tidal currents are required for various activities in the maritime industry, including marine power exploitation, operational forecasting, and coastal engineering. In addition to the conventional least squares harmonic method (LSHM) and numerical simulation, a machine learning algorithm has recently been developed and applied for predicting tidal currents in time series. Gaussian Process Regression (GPR) is one of the most important machine learning algorithms due to its ability to process time-series data. In this paper, we propose a combined one-dimensional LSHM and GPR method for predicting depth-averaged current velocity vectors according to time-series changes of water surface elevation due to tide. LSHM firstly decomposes water surface elevation time series into a number of possible constituents (i.e., amplitudes, phase-lags) according to known tidal frequencies. For each of the contributing constituents, GPR is then applied to produce the corresponding current generator using known tidal current as training data. The training data are generated by numerical simulations representing different coastal settings. The predicted tidal current is calculated according to the previously trained GPR model. As a comparison, we also generated tidal current predictions using a conventional two-dimensional LSHM. It is found that the tidal currents predicted using the sequence of computations proposed in this paper show comparable results with slightly better accuracy than conventional two-dimensional LSHM. The results from this study could provide an alternative method of tidal current prediction, minimising the need for the laborious re-runs of numerical simulations and lengthy current observations, as well as removing the requirement to carry out a two-dimensional analysis of the tidal current using LSHM.