摘要
通过构建采用外部输入的非线性自回归神经网络(NARX),利用1979年1月1日00时—2003年12月25日23时逐时的实测潮位数据和再分析气象数据结合调和分析预报结果搭建模型,对库克斯(Cuxhaven)港口2004—2018年中增水最大的两次风暴潮极端事件潮位进行预报和验证,同时对影响模型性能的参数进行量化分析。结果表明:在NARX神经网络延迟数为24 h时模型的精度最高,两次极端风暴潮验证下的R2分别为0.94和0.95,且在最高潮位时的误差分别为57.78 cm和26.55 cm。实验中模型在延迟数方面存在阈值,当延迟数为24 h时模型效果最佳,在延迟数达到阈值前模型的精度逐渐上升,超过该阈值后模型精度下降;输入时间数据序列的长短会影响模型的精度,序列越长模型精度越高,但影响效果会逐渐降低。
Using hourly tide level measurements from 00:00 on January 1,1979 to 23:00 on December 25,2003,meteorological reanalysis data,and reconciliation analysis of the forecast results,a storm surge forecasting model based on a nonlinear autoregressive model with exogenous inputs(NARX)neural network is conducted and validated in two storm surge extreme events with the largest water gain in Cuxhaven harbor from 2004 to mid-2018.The effects of the model parameters on the model's performance are quantitatively assessed.The results show that the model's accuracy is the highest when the NARX neural network's delay number is 24 hours,and the R2 is 0.94 and 0.95 for the two extreme storm surges,with errors of 57.78 cm and 26.55 cm at the highest tide level,respectively.The model's accuracy gradually increases before the delay number reaches the threshold of 24 hours,and gradually decreases after the delay number exceeds the threshold.The temporal duration of the input data also affects the model's accuracy,and longer input data series leads to higher accuracy of the model,but such relationship becomes weak as the temporal duration of the input data exceed a threshold.
作者
赵宏凯
迟万清
杨洁
周涛
ZHAO Hongkai;CHI Wanqing;YANG Jie;ZHOU Tao(First Institute of Oceanography,Ministry of Natural Resources,Qingdao 266061,China)
出处
《海洋预报》
CSCD
北大核心
2023年第3期11-18,共8页
Marine Forecasts
基金
国家重点研发计划(2022YFC2803800)。
关键词
NARX神经网络
风暴潮潮位
潮位预报
NARX neural network
storm tide elevation
tidal level prediction
