摘要
准确的入库日径流预测在水库优化调度中发挥着重要作用.为提高日径流预测精度,提出了基于小波包变换(WPT)并结合了白鲨优化(WSO)算法的门限循环控制单元(GRU)、长短期记忆神经网络(LSTM)、卷积神经网络(CNN)日径流时间序列预测模型,以云南省暮底河水库2018—2020年入库日径流时间序列预测为例对各模型进行检验.首先利用WPT将日径流时序数据分解为若干子序列分量;其次引入WSO对GRU、LSTM、CNN超参数进行调优,建立WPT-WSO-GRU、WPT-WSO-LSTM、WPT-WSO-CNN模型;最后利用所建立的模型对各子序列分量进行预测及加和重构,并构建WPT-GRU、WPT-LSTM、WPT-CNN及基于BP神经网络的WPT-WSO-BP、WPT-BP作对比分析模型.结果表明:WPT-WSO-GRU、WPT-WSO-LSTM、WPT-WSO-CNN模型对实例日径流预测的平均绝对百分比误差EMAP分别为3.67%、5.52%、8.98%,平均绝对误差EMA分别为0.120、0.155、0.329 m^(3)/s,确定性系数DC分别为0.996 2、0.995 7、0.974 0 s,预报合格率RQ分别为98.1%、96.4%、89.6%,预测效果优于对应未经WSO调优的WPT-GRU、WPT-LSTM、WPT-CNN模型及WPT-WSO-BP、WPT-BP模型,其中WPT-WSO-GRU模型具有更高的预测精度和更好的泛化能力,WPT-WSO-LSTM模型次之.WSO能有效调优GRU、LSTM、CNN超参数,提高GRU、LSTM、CNN预测性能.WPT-WSO-GRU、WPT-WSO-LSTM模型在入库日径流时间序列预测研究中具有较好的应用前景.
Accurate daily inflow prediction plays an important role in reservoir optimal operation.In order to improve the accuracy of daily runoff prediction,a threshold cycle control unit(GRU),a long and short term memory neural network(LSTM),and a convolutional neural network(CNN)daily runoff time series prediction models based on wavelet packet transform(WPT)and white shark optimization(WSO)algorithm are proposed.The models are tested by the daily inflow time series data of the Mudihe Reservoir in Yunnan Province from 2018 to 2020.Firstly,the WPT is used to decompose the daily runoff time series data into several subsequence components.Secondly,the WSO is introduced to adjust the super parameters of GRU,LSTM and CNN,and thus WPT-WSO-GRU,WPT-WSO-LSTM and WPT-WSO-CNN models are established.Finally,the established model is used to predict and reconstruct the sub sequence components,and the WPT-GRU,WPT-LSTM,WPT-CNN,and WPT-WSO-BP and WPT-BP models based on the BP neural network are constructed for comparison and analysis.The results show that the average absolute percentage error MAPE of WPT-WSO-GRU,WPT-WSO-LSTM and WPT-WSO-CNN models for daily runoff prediction of the case is 3.67%,5.52%and 8.98%respectively.The mean absolute error(MAE)is 0.120 m^(3)/s,0.155 m^(3)/s and 0.329 m^(3)/s respectively,and the determination coefficient(DC)is 0.9962,0.9957 and 0.9740 s respectively.The prediction qualification rate(QR)is 98.1%,96.4%and 89.6%respectively,which is better than that of WPT-GRU,WPT-LSTM WPT-CNN model and WPT-WSO-BP,WPT-BP models without WSO optimization.Among those models,the WPT-WSO-GRU model has higher prediction accuracy and better generalization ability,followed by WPT-WSO-LSTM model.WSO can effectively tune the super parameters of GRU,LSTM and CNN to improve the prediction performance.The WPT-WSO-GRU and WPT-WSO-LSTM models have a good application prospect in the prediction of the time series of daily inflow time series prediction research.
作者
陈金红
崔东文
CHEN Jinhong;CUI Dongwen(Yunnan Water Resources and Hydropower Investment Co.,Ltd.,Kunming 650051,China;Yunnan Province Wenshan Water Bureau,Wenshan 663000,China)
出处
《三峡大学学报（自然科学版）》
CAS
2023年第4期25-32,共8页
Journal of China Three Gorges University：Natural Sciences
基金
澜沧江非一致性径流演变规律及驱动机制研究(91547205)
云南省创新团队建设专项(2018HC024)
国家澜湄合作基金项目(2018-1177-02)。
关键词
日径流预测
门限循环控制单元
长短期记忆神经网络
卷积神经网络
白鲨优化算法
小波包变换
daily inflow forecast
gated recurrent unit
long-short term memory neural network
convolutional neural network
white shark optimization algorithm
wavelet packet transform
