| 摘要: |
| 洪涝灾害是世界主要自然灾害之一，优化洪水预报方案对防洪决策至关重要，然而传统水文模型存在参数多、调参受人为因素影响，泛化能力弱等问题。针对上述问题，本文提出基于改进的鲸鱼优化算法和长短期记忆网络构建自动优化参数的WOA-LSTM模型，并且通过优化神经网络结构进一步增强了该模型的稳定性和精确度，并且建立了不同预见期下的洪水预报模型来分析讨论神经网络结构与预报期之间的关系。以横锦水库流域1986-1997年洪水资料为例，其中以流域七个雨量站点的降雨以及横锦站水文资料为输入，不同预见期下洪水过程作为输出，以1986年-1993年作为模型的率定期，1994年-1997年作为模型的检验期，研究结果表明：（1）以峰现时差、确定性系数、径流深误差和洪峰流量误差作为评价指标，相比较于LSTM模型和新安江模型对检验期的模拟结果表明WOA-LSTM模型拥有更高的精度、预报结果更稳定。（2）结合置换特征值和SHAP法分析模型特征值重要性，增强了神经网络模型的可解释性。（3）通过改变神经网络结构一定程度避免由于预见期增加，数据关联性下降而导致的模型预报精度下降的问题，最终实验表明该模型在预见期1-6h下都可以满足横锦水库的洪水预报要求，可以为当地的防洪决策提供依据。 |
| 关键词: 洪水预报 LSTM 鲸鱼优化算法 深度学习 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金（41971137、U2240224、42001109），江苏省自然科学基金（BK20201102）和江西省科技计划项目（20213AAG01012、20212BBG71002、20222BCD46002）联合资助 |
|
| Research on reservoir flood forecasting method based on WOA-LSTM |
|
dingyiding1, fanhongxiang2, jiangmingliang3,4,4,5, xu li gang2, lvhaishen1
|
|
1.College of Hydrology and Water Resources, Hohai University;2.Nanjing Institute of Geography & Limnology Chinese Academy of Sciences;3.Nanjing Institute of Geography &4.amp;5.Limnology Chinese Academy of Sciences
|
| Abstract: |
| Flood disasters are one of the major natural disasters in the world. Optimizing flood forecasting solutions is crucial for flood management decisions. However, traditional hydrological models suffer from issues such as a high number of parameters, susceptibility to human interference during parameter calibration, and weak generalization ability. To address these challenges, this study proposes a WOA-LSTM model that integrates an improved Whale Optimization Algorithm and Long Short-Term Memory network to automatically optimize model parameters. Furthermore, the stability and accuracy of the model are further enhanced through the optimization of the neural network structure. Additionally, flood forecasting models with different lead times are established to analyze and discuss the relationship between the neural network structure and the forecast period. Taking the flood data of Hengjin Reservoir basin from 1986 to 1997 as an example, the rainfall data of seven stations and the hydrological data of Hengjin Station were used as input, the flood process under different forecast periods was used as output, 1986-1993 was used as the calibration period, and the validation period of the model was used from 1994 to 1997, and the results that: (1) Taking the peak-present time difference, deterministic coefficient, relative error in runoff depth and peak flow error as the evaluation index in contrast with that of the simulation results of the LSTM model and the XAJ model on the test period, the results show that the WOA-LSTM model has higher accuracy and more stable forecast results. (2) By displacing the eigenvalues, the SHAP method analyzes the importance of the model eigenvalues, which enhances the interpretability of the neural network model. (3) By appropriately changing the neural network structure, the problem of model prediction accuracy deteriorating due to the decrease of data correlation caused by the increase of the forecast period can be avoided to a certain extent, and finally the experimental results show that the model can meet the flood forecasting requirements of Hengjin Reservoir during the forecast period of 1-6h, and can provide a basis for local flood control decisions. |
| Key words: flood forecasting LSTM whale optimization algorithms deep learning |
