| 本文已被：浏览 77次 下载 19次 |
|
|
|
| 考虑记忆时间的LSTM在赣江流域径流预报中的应用 |
|
胡乐怡1, 蒋晓蕾1,2,3, 周嘉慧1, 欧阳芬4, 戴逸姝1, 章丽萍1, 付晓雷1,2,3
|
|
1.扬州大学水利科学与工程学院;2.水利部水文气象灾害机理与预警重点实验室，南京信息工程大学;3.河海大学水灾害防御全国重点实验室;4.南昌工程学院，水利与生态工程学院
|
|
| 摘要: |
| 在气候变化条件下，准确的径流预测对水资源的规划与管理十分重要。本文基于长短时记忆神经网络（LSTM）模型，采用赣江流域外洲、峡江以及栋背水文站的逐日流量以及CN05.1日降水数据构建了三个不同面积流域的径流预测模型，并通过设置不同情景分析：模型的有效预见期与不同流域平均产汇流时间之间的关系，有效预见期内LSTM径流预测模型精度与记忆时间之间的关系，不同长度的预见期与模型最佳记忆时间之间的关系，同时探讨LSTM径流预测所需的记忆时间与流域面积的关系。研究结果表明：（1）综合考虑降水和前期径流情景下的径流预测效果最好，当预见期为1天时，外洲、峡江、栋背站的纳什效率系数（NSE）分别可达0.98、0.96以及0.90。且其有效预见期与仅考虑降水信息的有效预见期相同，均与流域平均产汇流时间相近。（2）随着预见期的延长，不同情景下的预测精度均有不同程度的下降，其中仅考虑前期径流情景的下降率最大，说明降水信息较前期径流对径流预测效果的提升更重要。同时，随着流域面积的增加，相同预见期内径流预测精度均有所提升。（3）当预见期相同时，随记忆时间的延长，不同径流预测模型的预测精度均先上升至最高，接着具有下降趋势，最后逐渐趋于稳定。且在有效预见期内，随着预见期的延长，最佳记忆时间均有增大趋势，当达到最长的有效预见期时，对应的最佳记忆时间均为14天。此外，在赣江流域的模拟结果表明随着流域面积的增大，LSTM的最佳记忆时间减小。研究结果可为赣江流域的径流预报提供参考，同时有助于推求其他流域采用机器学习进行径流预测所需的最佳记忆时间。 |
| 关键词: LSTM 赣江流域 记忆时间 径流预测 预见期 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金资助项目(42371021,52109036)；河海大学水灾害防御全国重点实验室“一带一路”水与可持续发展科技基金面上项目(2022491111, 2021490611)；水利部水文气象灾害机理与预警重点实验室开放基金(HYMED202203, HYMED202210)；江苏省研究生科研与实践创新计划项目(KYCX23_3546, KYCX23_3549) |
|
| The application of LSTM considering time steps in runoff prediction of Ganjiang River Basin |
|
HU Leyi1, JIANG Xiaolei, ZHOU Jiahui1, OUYANG Fen, DAI Yishu1, ZHANG Liping1, FU Xiaolei1,2,3,4
|
|
1.College of Hydraulic Science and Engineering,Yangzhou University;2.Key Laboratory of Hydrometeorological Disaster Mechanism and Warning of Ministry of Water Resources, Nanjing University of Information Science &3.Technology;4.The National Key Laboratory of Water Disaster Prevention, Hohai University
|
| Abstract: |
| Accurate runoff forecasting is important for the water resources management and planning under climate change. We constructed the Long Short-Term Memory (LSTM) neural network model for runoff forecasting at three catchment areas (Waizhou, Xiajiang, and Dongbei) with different sizes in Ganjiang River basin. Based on the constructed LSTM, we analyzed the relationship between the effective lead time of the model and watershed average transit time of different basins, between the accuracy of the LSTM runoff prediction model and time steps during the effective lead time, between different lead time and the best time steps of the model, as well as the relationship between the time steps required for LSTM runoff prediction and the watershed area. The results show that (1) the best forecasted runoff can be obtained by considering both precipitation and antecedent runoff. The Nash-Sutcliffe efficiency coefficient (NSE) for the Waizhou, Xiajiang and Dongbei stations can reach 0.98, 0.96 and 0.90, respectively under the 1 day lead time. Moreover, the effective lead time is the same to the one which only considering the information of precipitation, and both are close to the watershed average transit time. (2) When extending the lead time, the runoff prediction accuracy under different scenarios decreased, and the largest decreasing rate is the scenario which only considering the antecedent runoff. This indicates that the precipitation is more important for runoff prediction. Additionally, the accuracy of runoff prediction increased when increasing the watershed area. (3) When fixing the lead time, the runoff prediction accuracy firstly rises, then decreases to gradually stabilize with the extension of the time steps. During the effective lead time, the best time steps increased when extending the lead time, and the best time steps are 14 days for the longest effective lead time. In addition, the results also show that the best time steps of LSTM decreased when increasing the catchment area in the Ganjiang River Basin. Thus, the results can be taken as a reference for runoff prediction in Ganjiang River Basin, and are helpful for obtaining the best time steps of machine learning or runoff prediction in other basins. |
| Key words: LSTM Ganjiang River Basin time steps runoff prediction lead time |
