A Spatiotemporal Machine Learning Framework for the Prediction of Metocean Conditions in the Gulf of Mexico: Application to Loop Current and Loop Current Eddy Forecasting

We are entering an exciting new era of data-driven weather prediction, where forecast models trained on historical data (including observations and reanalyses) offer an alternative to directly solving the governing equations of fluid dynamics. By capitalizing on a vast amount of available information – and capturing their inherent patterns that are not represented explicitly – such machine learning-based techniques have the potential to increase forecast accuracy, augmenting traditional physics-based equivalents. Here, we adapt and apply a promising machine learning framework – originally proposed by the present authors for regional prediction of ocean waves – to the operational forecasting of the Loop Current and Loop Current Eddies (LC/LCEs) in the Gulf of Mexico (GoM). The approach consists of using an attention-based long short-term memory recurrent neural network to learn the temporal patterns from a network of available observations, that is then combined with a random forest based spatial nowcasting model, trained on high-resolution regional reanalysis data, to develop a complete spatiotemporal prediction for the basin. Since machine learning approaches are typically physics-agnostic, an identical framework to that developed for the prediction of ocean waves can be used for the prediction of surface currents, with the only difference being the training datasets to which this is exposed. This is illustrated here using a period of three months of training data from October 2022 to December 2022, with the model driven by only three observation sites in the northern GoM. As such, it is unrealistic to expect performance for an unseen week in January 2023 to be equivalent to smaller/simpler domains with a more favorable quantity, quality and coverage/distribution of input observations but, despite these severe constraints, the ability of the model to forecast a plausible structure of the LC/LCE system is nonetheless impressive. The architecture of the MaLCOM framework allows for easy interrogation of the temporal and spatial behavior of the model which allows us to better unpick and explain its characteristics – thus providing a path to inform further enhancements. While still at an early stage of refinement, the extension of the framework from waves to currents demonstrates encouraging potential for a fundamentally different approach to the way that metocean data in general, and LC/LCE forecasts in particular, can be generated and used by the offshore energy sector, by directly leveraging sparse sensor networks as the basis for these predictions (further extending the value of the observations, when collected with this additional purpose in mind). Provided a suitable coverage, quality and quantity of observations are available, the advent of these very low cost, data-driven predictions – able to be run on-demand, in-house, using standard laptop or desktop computers – herald new opportunities for improving real-time decision-making to support offshore planning and workability.