耦合多变量筛选和多层LSTM的短期径流预测研究  

Study on Short-term Runoff Prediction Coupled with Multivariate Screening and Multilayer LSTM

作　　者：田伟 殷兆凯 董义阳 黄迪[1] 刘青 TIAN Wei;YIN Zhaokai;DONG Yiyang;HUANG Di;LIU Qing(Three Gorges Hubei Qingjiang Hydropower Development Co.,Ltd.,Yichang 443000,Hubei,China;Three Gorges Academy of Science and Technology,Beijing 101100,China;710 Research Institute of China State Shipbuilding Corporation,Wuhan 443100,Hubei,China)

机构地区：[1]三峡集团湖北清江水电开发有限责任公司,湖北宜昌443000 [2]三峡集团科学技术研究院,北京101100 [3]中国船舶集团710研究所,湖北武汉443100

出　　处：《水力发电》2025年第3期22-27,118,共7页Water Power

基　　金：国家重点研发计划项目(2022YFC3002702)。

摘　　要：径流预测影响因子筛选是流域来水预报研究过程中的关键环节。构建时序过程较复杂的水库短期径流预测模型时,能够输入模型的影响因子种类繁多,为了减少输入数据集维度,验证新的关键影响因子,以短期水库径流预测作为研究对象,建立不同尺度数据集的长短时记忆(LSTM)神经网络进行模型率定;再引入Fisher Score算法和熵权-TOPSIS法从16个相关水文气象、水库调度、发电调度等类型的常规影响因子中筛选出7个关键影响因子;然后以RMSE(均方根误差)为精度指标对3种LSTM模型的超参数进行优化;最后将各优化筛选后的参数、影响因子叠加到多层LSTM模型中对新的关键影响因子进行流量预测验证。研究发现,筛选影响因子后的LSTM模型率定效果更好,且新提出的上游水库发电计划执行偏差率这一影响因子能够进一步提升水库径流预测精度。The screening of influencing factors for runoff prediction is a key link in the process of basin water forecasting.When building a short-term reservoir runoff prediction model with complex time series process,there are a variety of influencing factors that can be input into the model.In order to reduce the dimensions of input dataset and verify the new key influencing factors,this paper takes the short-term reservoir runoff prediction as the research object,and establishes the long short-term memory(LSTM)neural network of different scale datasets for model calibration.Then the Fisher Score algorithm and entropy weight-TOPSIS method are introduced to select seven key influencing factors from sixteen conventional influencing factors related to hydrometeorology,reservoir scheduling and power generation scheduling,and the root mean square error(RMSE)is used as accuracy index to optimize the hyperparameters of the three LSTM models.Finally the parameters and influencing factors of each optimized screening are superimposed into the multi-layer LSTM model to verify the flow prediction of the new key influencing factors.It is found that the LSTM model after screening the influencing factors has a better calibration effect,and the newly proposed influencing factor of the deviation rate of the execution of upstream reservoir power generation plan can further improve the prediction accuracy of reservoir runoff.

关 键 词：长短时记忆 径流预测 Fisher Score算法 水库调度 发电计划执行偏差率 关键影响因子 

分 类 号：P333.1[天文地球—水文科学]