Spatiotemporal Prediction of Tidal Currents Using Gaussian Processes

Predicting fast tidal currents can be a challenging task. Unlike tidal water levels, currents can vary sharply over short distances. The classical approach of harmonic analysis can analyze data at point locations and there is a need for a method that can handle spatiotemporal data, as well as be robust to the uncertainty and noise inevitable in real-world measurements. In this work, we present a Bayesian machine learning (ML) approach to tackle the problem. The method is based on Gaussian processes, a nonparametric ML technique that uses a kernel function to capture structures in the data. A case study is performed using data from a validated numerical model simulating the tidal dynamics in the Pentland Firth region, UK. Several sampling strategies are explored and the case where measurement location is changed after every sampling is found to produce the lowest average error in the predictions. We show that the presented methodology using data from just a single moving data source can provide a better spatiotemporal description than traditional techniques using continuous data from a large number of locations. The work can be useful to developers of tidal energy farms, navigation, and other purposes.