摘要
为提高水库来水量时间序列预测精度,建立了小波包分解(WPD)-蝠鲼觅食优化(MRFO)算法-回声状态网络(ESN)相融合的时间序列预测模型,利用WPD将非平稳水库来水量时间序列分解为若干高频和低频时间序列,以便有效降低来水量时间序列的复杂性。在不同维度条件下选取8个典型函数对MRFO算法进行仿真测试,利用MRFO算法对ESN储备池规模、稀疏度等关键参数进行优化以提高网络训练效率。随后构建了WPD-MRFO-SEN模型和WPD-MRFO-SVM模型,并将这两个模型的预测结果和经经验模态分解(EMD)的EMD-MRFO-ESN模型和EMD-MRFO-SVM模型的结果作对比分析。利用云南省暮底河水库1956—2017年逐月来水量时间序列数据对上述4种模型的结果进行检验。结果表明:MRFO算法具有较好的寻优精度和全局搜索能力;WPD-MRFO-SEN模型对实例后10年120个月来水量时间序列预测的平均绝对百分比误差为2.23%,平均绝对误差为23.3万m3,均方根误差为35.8万m3,预测精度优于WPD-MRFO-SVM模型的,明显优于EMD-MRFO-ESN模型和EMD-MRFO-SVM模型的,具有较高的预测精度。WPD对水库来水量时间序列数据的分解效果优于EMD方法的。
In order to improve the prediction accuracy of the time series of reservoir inflow,a time series prediction model based on Wavelet Packet Decomposition(WPD)-Manta Ray Foraging Optimization(MRFO)algorithm-Echo State Network(ESN)was established.WPD was used to decompose the time series of non-stationary reservoir inflow into several high-frequency and low-frequency time series,so as to effectively reduce the complexity of the time series of inflow.Under the conditions of different dimensions,eight typical functions were selected to test the simulation results of MRFO algorithm.To improve the network training efficiency,MRFO algorithm was used to optimize the key parameters such as the scale and sparsity of the ESN reserve pool.Then WPD-MRFO-SEN model and WPD-MRFO-SVM model were constructed,and the prediction results of these two models were compared with the results of EMD-MRFO-ESN model and EMD-MRFO-SVM model.The results of the above four models were tested by using the monthly water volume time series data of Mudihe Reservoir in Yunnan from 1956 to 2017.The results show that MRFO algorithm has better optimizing precision and global searching ability;The average absolute percentage error,the average absolute error and the root mean square error of the prediction results of WPD-MRFO-SEN model are 2.23%,233000 m~3 and 358000 m~3,respectively;The prediction accuracy of WPD-MRFO-SEN model is better than those of WPD-MRFO-SVM model,EMD-MRFO-ESN model and EMD-MRFO-SVM model.The decomposing results for time series data of reservoir inflow by WPD is better than those by EMD.
作者
崔东文
CUI Dongwen(Wenshan Water Bureau of Yunnan,Wenshan 663000,China)
出处
《华北水利水电大学学报（自然科学版）》
北大核心
2022年第6期10-17,共8页
Journal of North China University of Water Resources and Electric Power：Natural Science Edition
关键词
来水量预测
小波包分解
蝠鲼觅食优化算法
回声状态网络
仿真测试
inflow forecast
Wavelet Packet Decomposition
Manta Ray Foraging Optimization algorithm
Echo State Network
simulation test
