[A-AS04] Machine Learning Techniques in Weather, Climate, Hydrology and Disease Predictions

[AAS04-11] Using neural network and parametric cyclone model for long-lead-time prediction of storm surge: A case study in the northeast of Taiwan

*Wei-Ting Chao¹, Chih-Chieh Young² (1.Center of Excellence for Ocean Engineering, National Taiwan Ocean University, Keelung, ROC., 2.Department of Marine Environmental Informatics, National Taiwan Ocean University, Keelung, ROC.)

Keywords:Storm surge, BPNN, parametric cyclone model, pressure, wind

Typhoon-induced storm surge is one of the most serious disasters on the earth. Irregular sea surface rise was driven by strong winds and atmospheric pressure disturbance. The Lifes and properties of coast residents’ often face these severe extreme nature hazards threats. Therefore, how to achieve a better prediction of storm surge is very important. In this study, a storm surge prediction model is proposed which is combining backward propagation neural network (BPNN) and parametric cyclone model. Some problems have been conquered through the advantages of the above mentioned two methods. First, the BPNN model could easily capture the nonlinear relationship between in-situ data and meteorological conditions. These kinds of data-driven tools also reduce a lot of computational resources waste compared to most full-physics coupled storm surge models. Second, the parametric cyclone models provide more effective information through incorporating into the BPNN model, which not only reduce input dimensions but also improve the training/testing performance. Finally, based on error tolerance and generalization capability of BPNN can be remedied the parametric cyclone models. The BPNN model for storm surge prediction could be improved through utilized the same input parameters (e.g., water level, local wind speed, and local pressure, etc.) from parametric cyclone models than the original typhoon information. The prediction performance of each parametric cyclone model is also examined by three statistical indices, i.e., the root-mean-square error (RMSE), the coefficient of correlation (CC), and the coefficient of efficiency (CE).