A multi-layer perceptron approach for accelerated wave forecasting in Lake Michigan

Abstract A machine learning framework based on a multi-layer perceptron (MLP) algorithm was established and applied to wave forecasting in Lake Michigan. The MLP model showed desirable performance in forecasting wave characteristics, including significant wave heights and peak wave periods, considering both wind and ice cover on wave generation. The structure of the MLP regressor was optimized by a cross-validated parameter search technique and consisted of two hidden layers with 300 neurons in each hidden layer. The MLP model was trained and validated using the wave simulations from a physics-based SWAN wave model for the period 2005–2014 and tested for wave prediction by using NOAA buoy data from 2015. Sensitivity tests on hyperparameters and regularization techniques were conducted to demonstrate the robustness of the model. The MLP model was computationally efficient and capable of predicting characteristic wave conditions with accuracy comparable to that of the SWAN model. It was demonstrated that this machine learning approach could forecast wave conditions in 1/20,000th to 1/10,000th of the computational time necessary to run the physics-based model. This magnitude of acceleration could enable efficient wave predictions of extremely large scales in time and space.