征服现有的单一模型的局限性,一些混合算法等数据预处理方法是利用数据驱动的模型,希望提高复杂水文时间序列数据的预测性能通过提取时变组件与降噪。这些基于预处理的混合模型已经应用于水文(
2]。混合模型的框架通常由“分解”,“预测”和“合奏”[
2,
6,
13]。最常用的数据预处理方法是小波分析(WA)用于非线性和非平稳的水文数据分解成多尺度组件(
13]。这些多尺度处理组件进一步作为输入用于黑盒模型在预测阶段,最后预测组件整体得到最终的预测。彭et al。
6]提出的混合模型通过使用经验可靠的流量预测的小波变换和安。他们展示了他们在单一模型提出了混合模型的效率。之后,吴et al。
11)利用两阶段混合模型,通过融合小波多分辨率分析(WMRA),马和其他数据预处理方法,奇异谱分析和安提高日常流的估计。他们提出的五个模型包括ANN-MA ANN-SSA1 ANN-SSA2, ANN-WMRA1, ANN-WMRA2和建议与MA模型分解执行比WMRA更好。改善小波分解方法已经得到更精确的混合结果由佤邦(
17]。然而,佤邦的问题,减少了性能,即。,年代elect我on of mother wavelet basis function, is still an open debate as the selection of mother wavelet is subjectively determined among many wavelet basis functions. The optimality of multiscale characteristics entirely depends on the choice of mother wavelet function as poorly selected mother wavelet function can lead to more uncertainty in time-scale components. To overcome this drawback, Huang et al. [
18)提出了一个纯粹的数据驱动的经验模态分解(EMD)方法。EMD的目标是将非线性和非平稳的数据自适应分解成许多振动组件称为固有模式分解(IMF)。许多研究已经进行EMD结合数据驱动模型(
15,
18- - - - - -
21]。特别是在水文,EMD和安用于风速和流量预测(
15,
20.]。阿加尼亚和Homaifar
21]发达EMD-based预测深层信念网络准确预测和预测标准化流流动指数(SSI)。他们的研究显示,他们提出的模型比现有标准方法SSI的预测精度的提高。然而,康et al。
22]表明EMD遭受模式混合问题最终影响分解的效率。为了处理这种模式混合问题,吴和挂
23)提出了一种改进的EMD先后引入高斯白噪声的信号,称为集成经验模态分解(EEMD)地址的问题频繁明显EMD的混合模式。后,EEMD有效地用作数据分解方法提取多尺度特征(
24- - - - - -
26]。Di et al。
2)提出了一个四级混合模型(基于EEMD分解)通过减少冗余噪声来提高预测精度和适当的数据得出结论,耦合分解EEMD方法与数据驱动的模型可以改善预测性能相比,现有的基于EMD的混合模型。江et al。
26)提出另一个两阶段混合方法与数据驱动模型耦合EEMD高速铁路客流预测估计每天的客流量。他们建议他们提出的混合模型是由会计更适合短期预测每天的变化比其他混合动力和单一的模型。然而,由于连续添加独立的高斯白噪声,EEMD性能的影响,通过EEMD算法减少了货币基金中提取的准确性。戴et al。
27)在他们的研究报道,EEMD基于混合模型没有执行适当由于独立噪声增加。
在这项研究中,提出了两种新颖的方法来提高水文时间序列的预测精度。两种模型有相同的布局除了去噪,在两种不同的方法被用来去除噪音从水文时间序列数据。在这两种模型,在分解阶段,EEMD的一个改良版本。CEEMDAN,用来发现振荡,即。,theh我ghto low frequencies in terms of IMF. At prediction stage, multimodels are used to accurately predict the extracted IMFs by considering the nature of IMFs instead of using only single stochastic model. The purpose of using multimodel is two-way: one is for accurately predicting the IMFs by considering the nature of IMFs and the other is to assess the performance of simple and complex models after reducing the complexity of hydrological time series data through decomposition. Predicted IMFs are added to get the final prediction of hydrological time series. The proposed three stages involve denoising (D-step), decomposition (Decompose-step), and component prediction (P-step), which are briefly described below:
P-step结果:我t一个l我c>首先提取和残余,采用三种方法预测更精确和near-to-reality结果。因此,一个传统的统计方法,即。,华宇电脑(p,d,问),与 two other nonlinear ML methods, i.e., GMDH-NN and RBFNN, are used to predict the IMFs and residuals of all four river inflows. The rivers inflow data of all four rivers are split: 70% for training set and 30% for testing set. The parameters and structure of models are estimated using 886 observations of rivers inflow. The validity of proposed and selected models is tested using 30% data of rivers inflow. After successful estimation of multimodels on each IMF and residual, the best method with minimum MRE, MAE, and MSE is selected for each IMF prediction. The testing results of proposed models with comparison to all other models for all four rivers’ inflow, i.e., Indus inflow, Jhelum inflow, Chenab inflow, and Kabul inflow, are presented in Table
3。拟议中的EMD-CEEMDAN-MM WA-CEEMDAN-MM模型预测结果充分证明的有效性与最低绝笔4例,美,MSE相比1 s [
8],2 s [
15,
17),和3 s [
2评价模型。然而,总的来说,该WA-CEEMDAN-MM模型达到最低的MSE相比其他EMD-CEEMDAN-MM提出的模型。最糟糕的预测模型是1 s,即。,华宇电脑,与out denoising and without decomposing the hydrological time series data with highest MSE. The predicted graphs of proposed model, i.e., EMD-CEEMDAN-MM, with comparison to 2-S models, i.e., with EMD based denoised for Indus and Jhelum river inflow, are shown in Figure
8和WA-CEEMDAN-MM相比,2 s模型,即。,佤邦去噪的基础上,如图所示
9。