Application of surrogate models in estimation of storm surge: A comparative assessment

Coastal storm surge hazard assessment has received increased attention due to major hurricane events in the last two decades. Robust hazard assessment requires accurate and efficient storm surge prediction models; however, existing numerical models are either high-fidelity but computationally demanding or low-fidelity but with real-time forecasting application. This dichotomy has prompted the development of surrogate models that leverage available synthetic/historical storm databases to build prediction models intended to better balance efficiency and accuracy. Despite numerous studies that have examined the application of various surrogate modeling methods in coastal response prediction, no study is available that compares all of the frequently used methods for storm surge prediction. Furthermore, in most of these studies, the discussion of the performance of these models is bounded to aggregated error metrics (e.g., RMSE, R). This study aims to provide a comprehensive framework for comparison and assessment of the performance of surrogate models based on Artificial Neural Network (ANN), Gaussian Process Regression (GPR) and Support Vector Regression (SVR) for predicting storm surge. The United States Army Corp of Engineers' North Atlantic Coast Comprehensive Study (NACCS) database is used for developing the models at representative coastal locations. In this study, the performance of the models is explored by investigating the stability of performance across training sample sizes, identifying systematic trends in errors, assessing performance in predicting large target response quantities, and characterizing the distribution of error. The results indicate that the performance of surrogate models may be improved by use of physically-motivated parameter scaling and that the selection of a surrogate modeling method should be informed by factors such as consistency in performance under a range of target surge elevations. In particular, the results suggest that accuracy of the tested surrogate models may be a function of the target surge elevation. Furthermore, results suggest that model performance should be assessed using factors beyond aggregate error metrics because such measures (particularly when used without modification to focus on risk-significant events) may give incomplete information about performance of surrogate models. (C) 2020 Elsevier B.V. All rights reserved.