Forecasting hurricane-forced significant wave heights using a long short-term memory network in the Caribbean Sea

Abstract

Abstract. A long short-term memory (LSTM) neural network is proposed to predict hurricane-forced significant wave heights (SWHs) in the Caribbean Sea (CS) based on a dataset of 20 CS, Gulf of Mexico, and western Atlantic hurricane events collected from 10 buoys from 2010–2020. SWH nowcasting and forecasting are initiated using LSTM on 0, 3, 6, 9, and 12 h horizons. Through examining study cases Hurricanes Dorian (2019), Sandy (2012), and Igor (2010), results illustrate that the model is well suited to forecast hurricane-forced wave heights much more rapidly at a significantly cheaper computational cost compared to numerical wave models, with much less required expertise. Forecasts are highly accurate with regards to observations. For example, Hurricane Dorian nowcasts had correlation (R), root mean square error (RMSE), and mean absolute percentage error (MAPE) values of 0.99, 0.16 m, and 2.6 %, respectively. Similarly, on the 3, 6, 9, and 12 h forecasts, results produced R (RMSE; MAPE) values of 0.95 (0.51 m; 7.99 %), 0.92 (0.74 m; 10.83 %), 0.85 (1 m; 13.13 %), and 0.84 (1.24 m; 14.82 %), respectively. In general, the model can provide accurate predictions within 12 h (R≥0.8) and errors can be maintained at under 1 m within 6 h of forecast lead time. However, the model also consistently overpredicted the maximum observed SWHs. From a comparison of LSTM with a third-generation wave model, Simulating Waves Nearshore (SWAN), it was determined that when using Hurricane Dorian as a case example, nowcasts were far more accurate with regards to the observations. This demonstrates that LSTM can be used to supplement, but perhaps not replace, computationally expensive numerical wave models for forecasting extreme wave heights. As such, addressing the fundamental problem of phase shifting and other errors in LSTM or other data-driven forecasting should receive greater scrutiny from Small Island Developing States. To improve models results, additional research should be geared towards improving single-point LSTM neural network training datasets by considering hurricane track and identifying the hurricane quadrant in which buoy observations are made.