Hybrid modeling of wind waves in estuary based on machine learning and SWAN

Abstract:

Numerical models solving the wave action balance equation have been widely used to simulate wind waves, a driver of coastal erosion hazard. In-situ measurements, albeit sparse, are crucial to the calibration and validation of numerical models. In this study, a novel hybrid model is developed by integrating the physics-based SWAN model with the machine learning (ML) algorithms to predict wind waves in a shallow estuary. Two ML methods, bagged regression tree (BRT) and artificial neural network (ANN), are employed. The significant wave height and peak wave period are predicted based on the measured wind, waves and water depth using the physics-based model (SWAN), data-driven (ANN) model, and hybrid models (BRT-SWAN and BRT-SWAN-ANN). The performances of different models are compared based on statistical measures of model predictive skills. The BRT algorithm identified that the wind direction and energy dissipation by bottom friction have the largest influence on the SWAN prediction of wave height and wave period, respectively. The integrated BRT-SWAN model reduces the root-mean square error in the wave height and wave period predictions by 22.9% and 64.1%, respectively, compared with the SWAN results at the measurement location. Moreover, it is found that the field measurement of wind waves can be replaced by the BRT-SWAN model results for training the ANN model and achieve similar accuracy. Thus, the BRT-SWAN-ANN model can be utilized to estimate long-term wave parameters when field observation is unavailable. This study shows that the hybrid modeling approach is a useful tool for rapid prediction of wind waves.