Elsevier

Ocean Engineering

Volume 215, 1 November 2020, 107812
Ocean Engineering

Genetic Programming for storm surge forecasting

https://doi.org/10.1016/j.oceaneng.2020.107812Get rights and content

Highlights

  • The paper proposes a new approach to using Genetic Programming (GP) to evolve models for storm surge forecasting.

  • GP can evolve more accurate models for storm surge forecasting than other existing machine learning methods.

  • The model evolved by GP is more interpretable than models evolved by other (black-box) methods such as neural networks.

  • GP can automatically select relevant features when evolving storm surge forecasting models.

Abstract

Storm surge is a genuine common fiasco coming from the ocean. Therefore, an exact forecast of surges is a vital assignment to dodge property misfortunes and to decrease a chance caused by tropical storm surge. Genetic Programming (GP) is an evolution-based model learning technique that can simultaneously find the functional form and the numeric coefficients for the model. Therefore, GP has been widely applied to build models for predictive problems. However, GP has seldom been applied to the problem of storm surge forecasting. In this paper, we propose a new method to use GP for evolving models for storm surge forecasting. Experimental results on datasets collected from the Tottori coast of Japan show that GP can evolve accurate storm surge forecasting models. Moreover, GP can automatically select relevant features when evolving storm surge forecasting models, and the models evolved by GP are interpretable.

Introduction

Storm surge, a rise in sea level due to low atmospheric pressure and strong winds, is a severe natural disaster coming from the sea. Storm surges are especially harmful when they happen at a high tide, combining the impacts of the surge and the tide (Lee, 2008). With over 600 million individuals living in low-lying coastal zones, coastal surges have devastating societal impacts. For example, the most remarkable recorded storm surge in the United States was produced by Storm Katrina in 2005, which created a storm surge 9 m tall within the town of Cove St. Louis, Mississippi. The overall misfortune as a result of Katrina is evaluated to surpass $100 billion (Muis et al., 2016). Whereas high-impact occasions will, without a doubt, happen within the future, the advancement of the forecast of storm surge is able to enormously reduce the misfortune of lives and possibly lessen the sum of property harm (Kim et al., 2016, 2018; Thuy et al., 2016).

A conventional approach to storm surge prediction is to utilize process-based numerical forecasting models, but this approach is often computationally expensive. An alternative way is to use (data driven) machine learning algorithms such as artificial neural networks (ANNs) for predicting surge levels using the features such as sea surface levels, winds, sea level pressures, and tropical cyclone characteristics (Kim et al., 2016). For instance, Lee (2008) proposed a neural network combined with a consonant examination to forecast storm surges in Suao Harbor station, Taiwan. The experimental results showed that storm surge is effectively anticipated using neural systems. De Oliveira at al (De Oliveira et al., 2009). proposed a neural network model to predict ocean level varieties related to meteorological occasions. The results showed that the model is able to capture the impacts of the climatic and maritime intuitions. You at al (You and Seo, 2009). combined neural networks and clustering algorithms to build a storm surge prediction model. The observed results demonstrated that the model can be used for effectively forecasting territorial storm surges. Kim at al (Kim et al., 2016) demonstrated the impact of using different feature sets on an artificial neural network-based after-runner surge forecast model. The experimental results indicated that the combination of surge level, sea-level pressure, drop of sea-level pressure, longitude and latitude of typhoon, sea surface level, wind speed and wind direction comprises the optimal feature sets for predicting the surge level with the lead time of 24 h in the area of Sakai Minato on the Tottori coast, Japan.

Genetic programming (GP) is an evolutionary procedure to create solutions in the form of computer programs (Koza, 1992). The ability of GP to learn the definition of a function itself from sample information makes it a great choice for symbolic regression. Therefore, GP has been widely used to build regression models for many real applications. For instance, based on predictions of stock-prices using GP, a possibly profitable trading strategy was proposed in (Kaboudan, 2000). Gaur and Deo (2008) proposed the use of GP to build a model for real-time wave forecasting. The estimates created by GP shown that it can be respected as a promising tool for future applications to ocean forecasts. Azamathulla and Ghani (2010) utilized GP to anticipate river pipeline scour, and the execution of GP was found to be more viable when compared with the results of regression equations and artificial neural systems modeling in anticipating the scour depth around pipelines.

Recently, GP has also been applied to forecast storm surge. Sahoo and Bhaskaran (2019) proposed to use GP for predicting storm surge and inundation characteristics resulting from tropical cyclones. Experiments used datasets collected from the coast of Odisha adjoining the Bay of Bengal. Experimental results showed that both ANNs and GP perform very well as evidenced from their validation with actual data. However, GP has not been investigated to build models for storm surge forecasting with a lead time. Moreover, the ability of GP to automatically select features and build interpretable models for storm surge forecasting has not been studied. Therefore, this paper will investigate the ability of GP to build models for forecasting storm surge levels.

This paper aims to develop accurate and interpretable models for storm surge forecasting based on GP approach. The proposed method is compared with other common machine learning-based storm surge forecasting models to answer the following research questions:

  • 1.

    Whether the GP-based storm surge forecasting models can be more accurate than machine learning-based storm surge forecasting models;

  • 2.

    Whether GP can automatically select relevant features when building storm surge forecasting models.

  • 3.

    Whether storm surge forecasting models evolved by GP are interpretable.

The rest of this paper is organized as follow. Section 2 gives a brief on GP and Machine Learning methods utilized in this paper. The proposed method and experiment design are shown in Sections 3 Genetic Programming for storm surge forecast, 4 Parameter settings. The results and analyses are displayed and examined in Section 5. Discussions are given in Section 6. Section 7 states conclusions and future work.

Section snippets

Genetic Programming

Genetic Programming (GP) is one of the foremost well known evolutionary algorithm methods motivated by natural selection to evolve solutions, as computer programs, to problems (Koza, 1992, 1994). Within the 1990s, GP was primarily applied to mainly simple problems since it was rather computationally expensive. Nowadays, with the exponential development in CPU control and the changes of GP frameworks, it has broadly been utilized to solve various real-world problems. The applications of GP cover

Problem statement

In the current research, we used the remotely transmitted hourly recorded meteorological and hydrodynamic data from the observation stations on the Tottori coast of Japan, which are exactly the same Kim et al. (2016). There are three types of features used to predict storm surge level:

  • meteorological parameters: wind speed (WS) (m/s), wind direction (WD) (degree), sea-level pressure (SLP) (hPa), and drop rate of sea-level pressure (DSLP) from average sea-level pressure at five sites. These

Parameter settings

In all experiments, we used the implementation of GP in the ECJ library (A Java-based Evolutionary Computation Research System) (Luke et al., 2006). Table 1 shows the parameters of our GP systems. Thirty independent runs of GP were performed for each experiment, yielding diverse solutions at each run. These independent runs have different seeds generated by the computer pseudo-random number generator. We also utilized elitism in our evolutionary process, which copied the best individual of the

Empirical evaluation of the proposed method

In this experiment of storm surge forecast, it was found that the difference in performance between GP with the number of generations was significant, as seen in Fig. 2(a) and (b), and 2(c). We choose values of the number of generations in [50, 500].

First, the correlation coefficient of surge forecast with 5h-lead time is examined. In Fig. 2(a), GP achieves the best performance with the maximum number of generations of 200. It seems that, in this case, the more number of generation number leads

Discussion on accuracy

To evaluate our evolved GP models in detail on the 5h-lead time forecasts compared to the other five learning methods, we calculated two records of the correlation coefficient (CC) and the normalized root mean square error (NRMSE) in rate as appeared in Fig. 6(a). It is observed that the value of CC in all of the models ranges from 0.93 to 0.99, while that of NRMSE is from 5% to 12%. Among these models, the one evolved with GP has smallest error and the largest correlation coefficient.

Conclusions and future work

This paper has proposed a method to use Genetic Programming (GP) for building storm surge forecasting models. Meteorological, hydrodynamic and typhoon-featured parameters were taken into Genetic Programming to build models forecasting storm surge levels with the lead times of 5h, 12h and 24h. The proposed method was evaluated on the datasets collected from the observation stations on the Tottori coast. The experiments compared Genetic Programming with other common machine learning methods on

CRediT authorship contribution statement

Nguyen Thi Hien: Methodology, Writing - original draft, Visualization. Cao Truong Tran: Methodology, Writing - original draft, Writing - review & editing, Formal analysis. Xuan Hoai Nguyen: Supervision, Validation, Writing - review & editing. Sooyoul Kim: Supervision, Validation, Writing - review & editing. Vu Dinh Phai: Data curation. Nguyen Ba Thuy: Resources, Funding acquisition. Ngo Van Manh: Resources, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

The research presented in this paper was funded by National Science and Technology Program to respond to climate change, manage natural resources and the environment in the period 2016–2020 in the project titled “Applications of Artificial Intelligence for Forecasting Hydro-Meteorological Anomalies in Vietnam under the Context of Climate Change”, Grant Number: BĐKH.34/16-20.

References (26)

  • S. Gaur et al.

    Real-time wave forecasting using genetic programming

    Ocean. Eng.

    (2008)
  • S. Kim et al.

    A real-time forecast model using artificial neural network for after-runner storm surges on the tottori coast, Japan

    Ocean. Eng.

    (2016)
  • Y. Song et al.

    An efficient instance selection algorithm for k nearest neighbor regression

    Neurocomputing

    (2017)
  • H.M. Azamathulla et al.

    Genetic programming to predict river pipeline scour

    J. Pipeline Syst. Eng. Pract.

    (2010)
  • R.M. Balabin et al.

    Support vector machine regression (svr/ls-svm)–an alternative to neural networks (ann) for analytical chemistry? comparison of nonlinear methods on near infrared (nir) spectroscopy data

    Analyst

    (2011)
  • C.W. Dawson et al.

    Hydrological modelling using artificial neural networks

    Prog. Phys. Geogr.: Earth Environ.

    (2001)
  • M.M. De Oliveira et al.

    Neural network model to predict a storm surge

    J. Appl. Meteorol. Climatol.

    (2009)
  • N. de Vos et al.

    Constraints of artificial neural networks for rainfall-runoff modelling: trade-offs in hydrological state representation and model evaluation

    Hydrol. Earth Syst. Sci.

    (2005)
  • J. Demšar

    Statistical comparisons of classifiers over multiple data sets

    J. Mach. Learn. Res.

    (2006)
  • M. Hall et al.

    The weka data mining software: an update

    ACM SIGKDD Explorations Newslett.

    (2009)
  • J. Han et al.

    Data Mining: Concepts and Techniques

    (2011)
  • M.A. Kaboudan

    Genetic programming prediction of stock prices

    Comput. Econ.

    (2000)
  • S.-W. Kim et al.

    A surrogate modeling for storm surge prediction using an artificial neural network

    J. Coast Res.

    (2018)
  • Cited by (0)

    1

    Also at AI Academy Vietnam, 489 Hoang Quoc Viet St, Hanoi.

    View full text