| 摘要: |
| 长短期记忆网络（LSTM）模型要求输入资料时序连续，选择合适的输入因子及其输入方式对于提高LSTM在洪水预报中的应用效果具有重要意义。本文针对洪水场次时序上不连续的特点，构建了洪水场次拼接的固定时间步长的输入方式1、洪水场次分开考虑的动态时间步长的输入方式2、洪水场次分开考虑的结合固定时间步长和动态时间步长的输入方式3，分别以降雨、径流、降雨和径流为输入因子结合3种输入方式设置了7种输入方案，比较了不同输入方案下的LSTM模型在建阳流域的应用效果。结果表明：（1）选取降雨及径流作为输入因子的LSTM模型（方案3、6和7）洪水计算结果优于仅以降雨（方案2和5）或仅以径流（方案1和4）作为输入因子的设置。（2）当预见期为1~2h时，方案3、6、7的预测结果无显著差距；预见期为3~5h时，结合固定步长和动态步长设置的方案7洪水预报结果最优。（3）对方案7分别采用预见期1~3h的多个模型和1h的单个模型进行滚动预报，多模型的方式在3h预见期内对洪峰的预测精度更高。研究成果可为LSTM洪水预报模型的输入因子选择和时间步长设置提供参考，结合固定步长和动态步长的设置可提高较长预见期下的洪水预报精度。 |
| 关键词: 长短时记忆神经网络 洪水预报 资料拼接 步长设置 建阳流域 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Comparison study on the application performance of LSTM flood forecasting model under different input methods |
|
Qin Rui, Zhang Xiaoqin
|
|
College of Hydrology and Water Resources,Hohai University
|
| Abstract: |
| The Long Short-Term Memory (LSTM) model requires continuous input data in time series. Choosing appropriate input factors and input methods is of great significance for improving the application performance of LSTM in flood forecasting. In this study, three input methods are set to consider the discontinuous temporal characteristics of flood events including a fixed time step input method 1 that concatenates flood events, a dynamic time step input method 2 that considers floods separately, and an input method 3 that combines fixed and dynamic time steps by separating treatment of flood events; seven input schemes are designed by considering the above input methods with rainfall, streamflow, and both rainfall and streamflow as input factors respectively; the application performance of LSTM model with different input schemes is compared in the Jianyang River Basin. The results show that: (1) The LSTM model with rainfall and streamflow as input factors (schemes 3, 6, and 7) yields better flood calculation results than those with only rainfall (schemes 2 and 5) or only streamflow (schemes 1 and 4) as input factors. (2) When the lead times are 1-2 hours, there is no significant difference in the prediction results of schemes 3, 6, and 7; when the lead times are 3-5 hours, the scheme 7 that combines fixed step size and dynamic step size settings has the best performance. (3) The scheme 7 is to use multiple models with the lead times of 1-3 hours and a single model with the lead time of 1 hour for rolling forecasting, the multi model approach has higher accuracy in predicting flood peaks within the lead time of 3 hours. The research results can provide reference for the selection of input factors and time step setting of LSTM flood forecasting model, combining the setting of fixed time step and dynamic time step can improve the accuracy of flood forecasting under longer lead times. |
| Key words: Long short-term memory neural network flood forecasting data splicing step size setting Jianyang River Basin |
