基于集合经验模态分解与BP组合模型的短期余水位预测

Short-term residual water level prediction based on EEMD-BP neural network combination model

水位在忽略观测误差的前提下,可分解为潮位和余水位,后者具有较强的空间相关性以及非平稳特征,是影响水位预报精度的主要因素。港口工程、航运计划编制等方面对实时高精度水位预报具有重要需求,这对余水位预报模型构建提出了更高要求。另外,利用高精度余水位预报模型可减少验潮站布设数量。针对余水位短期预测模型精度不高的现状,本文对余水位进行集合经验模态(EEMD)分解,获得余水位在时间序列上的本征模函数(IMF);使用快速傅立叶变换(FFT)分析各本征模函数的频谱特征;再利用BP神经网络对各个本征模函数进行训练,预测了未来6 h、12 h、24 h的余水位值。对哥伦比亚河下游河口处的3组典型验潮站的余水位数据的预测结果表明,在未来6 h、12 h内的余水位的预测精度达到厘米级,在24 h内接近厘米级,证明了该组合模型在余水位短期预测方面的可行性。

On the premise of ignoring observation errors,the water level can be decomposed into tidal water level and residual water level,and the latter has strong spatial correlation and non-stationary characteristics,which is the main factor affecting the accuracy of water level prediction.Port engineering,shipping planning and other aspects have important requirements for real-time and high-precision tidal prediction,which puts forward higher requirements for residual water level prediction model.In addition,the number of tide stations can be reduced by using high precision residual water level prediction model.In view of the low accuracy of the short-term residual water level prediction model,this paper decomposes the residual water level with the Ensemble Empirical Mode Decomposition(EEMD)to obtain the Intrinsic Mode Function(IMF)of residual water level in time series.Fast Fourier Transform(FFT)was used to analyze the characteristics of Intrinsic Mode Function.BP-neural network was used to train each Intrinsic Mode Function,and residual water level values at 6 h,12 h and 24 h in the future were predicted.The residual water level data of three groups of typical tide stations in the lower reaches of the Columbia River Estuary were tested and verified.It is proved that the accuracy of the combined model reached the centimeter level within 6 h and 12 h,and was close to the centimeter level within 24 h,which proves the feasibility of the combined model in short-term residual water level prediction.