| 摘要: |
| 针对水文资料缺乏流域机器学习模型建模困难的问题，本文提出了基于长短期记忆神经网络（long-short term memory neural network，LSTM）的区域化洪水预报方法。对水文气候相似区内各流域的水文及地形地貌特征数据进行归一化处理，以消除局地因素的影响，从而构建相似区内建模统一数据集，扩大样本数量，为建立乏资料流域洪水预报模型提供了可能。选择胶东半岛作为研究区进行了应用研究。为验证区域化模型在不同场景中的应用效果，设计了预报流域数据不参与建模，而仅根据区域内其他流域资料建模，以及预报流域的部分数据参与建模两种情景；此外，选取仅根据预报流域数据训练的“单流域模型”作为基准模型进行对比分析。结果表明，对本次研究的水文资料短缺流域，两种区域化模型均取得了较好效果，且都优于单流域模型。相较而言，考虑了预报流域数据的区域化模型精度更高，说明在区域化LSTM构建中融入预报流域的数据，可进一步提升区域化模型的精度。研究成果可为乏资料地区的洪水预报提供参考。 |
| 关键词: 长短期记忆网络 洪水预报 区域化模型 水文气候相似区 乏资料流域 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金项目（52379007，42371045）、浙江省水利厅科技计划项目（RC2214）和水利部重大科技项目（SKR-2022032）联合资助。 |
|
| Research on regional long-short term memory neural network (LSTM) flood forecasting model |
|
Ye Kejia,Liang Zhongmin,Chen Hongyu,Qian Mingkai,Hu Yiming,Wang Jun,Li Binquan
|
|
College of Hydrology and Water Resources, Hohai University
|
| Abstract: |
| Aiming at the difficulty of hydrological modeling in the catchments with few hydrological data, a regional flood forecasting method based on long-short term memory neural network (LSTM) is proposed in this paper. Through scalarizing the data of topographic and geomorphological factors of each catchment in the same hydroclimatic similarity zone, this method succeeds in eliminating the influence of local factors. On this basis, a unified modeling dataset in a similarity zone is constructed and the sample size is effectively expanded, which provides a possibility for the establishment of flood forecasting models in the data-sparse catchments. Jiaodong Peninsula is selected as the study area in this study. In order to verify the application effect of the regional model in different scenarios, two regional modeling schemes are designed in this paper. To be specific, the first scheme is to construct a regional model based on the data of other basins in the similarity zone, without the participation of the data of the forecast basin. The second scheme is to construct a regional model with the data both from the forecast basin and other basins in the same region. In addition, the “single-basin model”, which is trained only with forecast basin’s data, is selected as the benchmark model. The results indicate that, for data-scarce basins in this study, both regional models exhibit high accuracy and are superior to single-basin model. Comparatively, the regional model that incorporates data from the forecast basin outperform the other regional model, suggesting that the accuracy of the regional model can be further improved if the data of the forecast basin is incorporated into model construction. The study can provide a reference for flood forecasting in data-scarce catchments. |
| Key words: long-short term memory neural network flood forecasting regional model hydroclimatic similarity zone data-scarce catchments |
