Prediction of Wave Overtopping from Coastal Defence Structures using Neural Networks and Ensemble Machine Learning Approach

Abstract:

Robust prediction of wave runup and overtopping is essential for design of coastal infrastructures and ensuring safety of population and assets in coastal zones (Abolfathi et al., 2016). Existing wave overtopping predictions are based on empirical formulae from limited physical and numerical models. Data availability through sensors, remote sensing and advances in computational power enabled application of advanced data-driven methods to predict coastal processes. CLASH is a large dataset of physical, numerical and field-based measurements of wave overtopping from various types of un-defended and defended shorelines. Previous ML research on overtopping focused on the application of Artificial Neural Networks (e.g., Zanuttigh et al., 2016; Formentin et al., 2017) following similar architecture, feature selection and validation strategies.

This paper investigates ML approaches for prediction of the resilience of critical coastal infrastructures to wave overtopping using CLASH database. Robust data analysis and automated feature selection are described to ensure modelling results are unbiased. Bayesian Optimisation is performed for hyperparameter tuning, showing improvements compared to grid search techniques. A comprehensive analysis of the performance of ANN, Random Forest and Support Vector Machines (SVM) models are presented and the trade-offs of each method is discussed.

The results show that ANN models can benefit from the use of different layer types, such as Batch Normalisation. Furthermore, Random Forest and SVM are shown to be good competitors to ANN for prediction of wave overtopping from coastal defences. The promising potential of ensemble models is demonstrated by combining the results of the different ML methods into a single predictor.

Reference

Abolfathi, et al., 2016. doi.org/10.1016/j.oceaneng.2015.12.016

Formentin, S.M., et al., 2017. doi.org/10.1142/S0578563417500061

Zanuttigh, B., et al., 2016. doi.org/10.1016/j.oceaneng.2016.09.032