摘要
径流非平稳性及超参数率定容易陷入“维数灾”的问题限制了神经网络(Long Short-Term Memory,LSTM)在中长期径流预报中的应用,针对上述问题,采取“分解-预测-重构”的思路,在数据预处理阶段采用变分模态分解(Variational Mode Decomposition,VMD)进行分解重构,耦合PSO(Particle Swarm Optimization)、SSA(Sparrow Search Algorithm)优化算法长短期记忆神经网络(LSTM),构建VMD-PSO-LSTM、VMD-SSA-LSTM、VMD-LSTM和LSTM四类模型进行对比试验研究,并探讨时滞步长对预报精度的影响。将模型应用于金沙江流域石鼓水文站月平均径流预报:(1)VMD-PSO-LSTM预报精度略高于VMD-SSA-LSTM,其次是VMD-LSTM,最后是LSTM。VMD-PSO-LSTM预报成果的MAE、MAPE、NSE、QR、RMSE分别为160.4、0.135、0.949、0.811、252.8;VMD-PSO-LSTM预报成果评价指标MAE、MAPE和RMSE分别比LSTM模型减小53.92%、50.67%、55.86%;NSE、QR分别比LSTM模型提升28.34%、84.48%;(2)在一定范围内,时滞步长越大,MAE、MAPE和RMSE增大,NSE、QR下降。【结论】对径流时间序列预处理,能够降低时间序列的非平稳性;加入优化算法能够在超参数空间中寻找到较优超参数组合,使得LSTM拟合能力较强,同时保持一定的泛化能力,从而提高预报精度。此外,时滞步长是影响预报精度的重要因素,加入优化算法能降低模型对时滞步长的敏感性。
The non-stationarity of runoff and the tendency of hyperparameter calibration to fall into the"curse of dimensionality"limit the application of LSTM(Long Short Term Memory)in medium to long-term runoff forecasting.To solve this problem,the idea of"decomposition prediction reconstruction"is adopted,and the variational mode decomposition(VMD)is used to decompose the signal in the data preprocessing stage,Long-short-term memory neural network(LSTM)coupled with optimization algorithms,and four types of models of VMD-PSO-LSTM,VMD-SSA-LSTM,VMD-LSTM and single LSTM are constructed.The model is applied to the monthly average runoff forecast of Shigu Hydrological Station in the Jinsha River basin.The result show that:1)The prediction accuracy of VMD-PSO-LSTM is slightly better than that of VMD-SSA-LSTM,followed by VMD-LSTM,followed by the single LSTM.The Mean absolute error,mean absolute percentage error,Nash-Sutcliffe efficiency coefficient qualification rate,and root mean square error of VMD-PSO-LSTM prediction result are 160.4,0.135,0.949,0.811,and 252.8,respectively;The evaluation indicators MAE,MAPE,and RMSE of VMD-PSO-LSTM prediction result are reduced by 53.92%,50.67%,and 55.86%respectively compared to the single LSTM model;NSE and QR improved by 28.34%and 84.48%respectively compared to the single LSTM model;2)Within a certain range,the larger the Time lag step size are,the larger the MAE,MAPE,and RMSE will be,while the less NSE and QR will be.[Conclusion]Conclusion:Data preprocessing of runoff time series can reduce the non-stationary of the time series.Optimization algorithms can find optimal hyperparameter combinations in the hyperparameter space,allowing LSTM to maintain a certain degree of generalization ability while possessing high fitting ability,thereby improving prediction accuracy.In addition,the time lag step size is an important factor affecting prediction accuracy,and adding optimization algorithms can reduce the sensitivity of the model to time lag step size.The Mixture model proposed in this paper provides a new method for monthly runoff prediction.
作者
袁旦
谭尧耕
朱艳霞
高超
刘珂
董宁澎
YUAN Dan;TAN Yaogeng;ZHU Yanxia;GAO Chao;LIU Ke;DONG Ningpeng(State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation,Tianjin University,Tianjin300350,China;China Institute of Water Resources and Hydropower Research,Beijing 100080,China;Information Center of Ministry of Water Resources of the People′s Republic of China,Beijing 100053,China;Kunming Engineering Corporation Limited,Yunnan 650041,Kunming,China)
出处
《水利水电技术（中英文）》
北大核心
2024年第S01期28-38,共11页
Water Resources and Hydropower Engineering
基金
中国电力建设股份有限公司科技项目(DJ-HXGG-2021-04)
云南省重点研发计划项目(202203AA080010)
