A spatiotemporal machine learning framework for the prediction of metocean conditions in the Gulf of Mexico

Machine learning techniques offer the potential to revolutionize the provision of metocean forecasts critical to the safe and successful operation of offshore infrastructure, leveraging the asset-level accuracy of point-based observations in conjunction with the benefits of the extended coverage (both temporally and spatially) of numerical modelling and satellite remote sensing data. Here, we adapt and apply a promising framework – originally proposed by the present authors for the prediction of wave conditions on the European North West Shelf – to the waters of the Gulf of Mexico. 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 buoy observations, that is then combined with a random forest based spatial nowcasting model, trained on reanalysis data, to develop a complete framework for spatiotemporal prediction for the basin. By way of demonstration, the new method is applied for the short-range prediction of wave conditions up to 12 hours ahead, using in-situ wave observations from the sparse network of National Data Buoy Center locations as an input, with the corresponding spatial mapping learned from the physics-based Met Office WAVEWATCH III global wave hindcast. The full spatiotemporal forecast system is assessed using independent measurements in the vicinity of the Louisiana Offshore Oil Port, previously unseen by the machine learning model. Results show that accurate real-time, rapidly updating wave predictions are possible, available at a fraction of the computational cost of traditional physics-based methods. The success of the approach, combined with the flexibility of the framework, further suggest its utility in related metocean challenges. While still at an early stage of development into a fully relocatable capability, it is intended that this contribution provides a foundation to stimulate a series of subsequent efforts to help support improved offshore planning and workability – including (but not limited to) applications linked with better resolving spatial variability across renewable energy sites, predicting ocean current regimes in the proximity of oil & gas platforms, as well as informing adaptive sampling strategies conducted by autonomous vessels – where the adoption of such a machine learning approach, that can be run on a laptop computer, having the potential to revolutionize data-driven decision-making by the industry.