基于递归神经网络的风暴潮增水预测

PDF下载HTML

Title:

Prediction of storm surge based on recurrent neural network

作者:

雷森¹, 史振威¹, 石天阳¹, 高松², 李亚茹², 钟山²

1. 北京航空航天大学 宇航学院图像处理中心, 北京 100191;
2. 国家海洋局 北海预报中心, 山东 青岛 266000

Author(s):

LEI Sen¹, SHI Zhenwei¹, SHI Tianyang¹, GAO Song², LI Yaru², ZHONG Shan²

1. Image Processing Center, School of Astronautics, Beihang University, Beijing 100191, China;
2. Beihai Forecast Center of State Oceanic Administration, Qingdao 266000, China

关键词:

风暴潮增水; 预测; 数值预报; 机器学习; 静态数据; 时序特性; BP神经网络; 递归神经网络

Keywords:

storm surge; prediction; numerical forecast; machine learning; static data; temporal properties; BP neural networks; recurrent neural network

分类号:

TP751

DOI:

10.11992/tis.201706015

摘要:

风暴潮增水的准确预测能极大地减少人员伤害和经济损失，具有重要的实用价值。传统的风暴潮预报方法主要包括经验和数值预报，很难建立起相对准确的模型。现有的基于机器学习风暴潮预报方法大都只提取出静态数据间的关系，并没有充分挖掘出风暴潮数据背后的时序关联特性。文中提出了一种基于递归神经网络的风暴潮增水预测方法。本文对风暴潮时序数据进行特定的处理，并设计合适结构的递归神经网络，从而完成时序数据的预测。相较于传统的BP神经网络，递归神经网络能更好地应对时序数据的预测问题。将该方法用于潍坊水站的增水预测中，结果表明，相对于BP神经网络，递归神经网络能得到更好的预测结果，误差更小。

Abstract:

Accurately forecasting storm surges can greatly reduce personnel injuries and economic losses, and so has great practical value. Traditional methods for predicting storm surge mainly involve experience and numerical forecasting, which makes it very hard to establish accurate models. Most of today’s storm surge forecast methods based on machine learning only extract the relationships among static data and fail to identify the relevant time series properties of these data. In this paper, we propose a storm surge forecast method based on the recurrent neural network. The storm surge data is rearranged with particular treatments, and an appropriate recurrent neural network is designed to perform the prediction of the time series. Compared with traditional BP neural networks, the recurrent neural network can better forecast time series data. In this study, we used a recurrent neural network to predict surges at the Weifang gauge station. The results show that the recurrent neural network produces a better prediction with a smaller error than the BP neural network.