Prediction of Significant Wave Heights by an Ensemble of Neural Networks

Resumo

Due to the chaotic behaviour of the differential equations which model the problem of predicting ocean waves variables, a well-known strategy to overcome the difficulties is to run several simulations, varying the initial condition, and averaging the result of each, creating an ensemble. In recent years, with the increase in available data and computational power, machine learning algorithms have been applied as surrogates to traditional numerical models, yielding comparative or better results. This work presents a methodology to create an ensemble of different artificial neural network architectures, namely, MLP, RNN, LSTM, CNN, and a hybrid CNN-LSTM, aiming to predict significant wave height at five different locations on the Brazilian coast. The networks are trained using NOAA’s numerical reforecast data and target the residual between observational data and the numerical model output. A new strategy to create the training and target datasets is demonstrated. Results show that the framework is capable of producing highly efficient forecasts, with an average accuracy of 80%, and a significant reduction in computational cost.