ABSTRACT

This study investigates the overall performance of non-linear regression models based either on DNN or on RNN for the fast predictions of time-varying storm surge heights. We selected Tokyo-Bay as a case study site and conducted numerical simulations of storm surges induced by 151 recorded typhoons that passed around Japan from 1951 to 2017. Obtained time-series data of the storm surge heights at selected 10 target points were used as outcome variables of the non-linear regression model. The corresponding predictor variables are obtained from the time series of typhoon data and the recent time history of the storm surges height at the same target point. This study then performed a parametric study to explore the optimum model setups of: (i) types of optimizer; (ii) the number of hidden layers (iii) the number of nodes of each hidden layer; and (iv) the duration time and the lead time of the time-series data of the typhoon and the recent storm surge height. Besides the optimum conditions of the present regression model, it was found that RNN showed clearly better predictive skills than DNN, and the recent history of the storm surge did not significantly improve the predictive performance of the present regression model.

抽象的

这项研究调查了基于DNN或RNN的非线性回归模型的整体性能，以快速预测随时间变化的风暴潮高度。我们选择东京湾作为案例研究地点，并对1951年至2017年在日本各地传播的151次台风诱发的风暴潮进行了数值模拟。以选定的10个目标点处的风暴潮高度的时间序列数据作为结果非线性回归模型的变量。相应的预测变量是从台风数据的时间序列和同一目标点的风暴潮高度的最近时间历史中获得的。然后，该研究进行了参数研究，以探索以下方面的最佳模型设置：（i）类型的优化器；（ii）隐藏层的数量（iii）每个隐藏层的节点数量；（iv）台风时间序列数据的持续时间和提前时间以及最近的风暴潮高度。除了当前回归模型的最佳条件之外，还发现RNN明显比DNN表现出更好的预测技能，并且风暴潮的近期历史并没有显着改善当前回归模型的预测性能。