基于SAO优化的LSTM模型在北流河的径流预报研究

郑凯丰; 马兴华; 崔国韬; 陶昌弟

doi:10.13928/j.cnki.wrahe.2026.02.006

水利水电技术（中英文） ›› 2026, Vol. 57 ›› Issue (2) : 83 -94. DOI: 10.13928/j.cnki.wrahe.2026.02.006

复杂灾害链与水旱巨灾风险评估专栏

基于SAO优化的LSTM模型在北流河的径流预报研究

郑凯丰 ¹^,² ,
马兴华 ³^,⁴ ,
崔国韬 ¹^,² ,
陶昌弟 ⁵

作者信息 +

Runoff forecasting in Beiliu River based on SAO-optimized LSTM model

Kaifeng ZHENG ¹^,² ,
Xinghua MA ³^,⁴ ,
Guotao CUI ¹^,² ,
Changdi TAO ⁵

Author information +

文章历史 +

PDF (1173K)

摘要

【目的】受气候变化与人类活动影响，极端径流事件所引发的灾害频发，基于深度学习模型长短时记忆神经网络(Long Short-Term Memory, LSTM)的径流预报广泛应用，但其仍需提升准确性与可解释性。【方法】开发了基于雪消融优化算法(Snow Ablation Optimizer, SAO)与LSTM的组合模型进行径流预报。选取流域水文气象特征(径流量、降雨量、气温)和基于皮尔逊相关系数(Pearson Correlation Coefficient,PCC)筛选的六项大尺度气候因子指数作为模型输入，并与反向传播神经网络(Back Propagation, BP)比较，最后使用Shapley值法分析输入特征重要性与贡献度。【结果】结果显示：SAO-LSTM模型平均绝对百分比误差(MAPE)为0.26,决定系数(R²)为0.80,Nash-Sutcliffe效率系数(NSE)为0.80,显著优于LSTM模型和BP模型，且表现出良好的泛化能力；Shapley解译结果显示降雨是关键驱动因子，大尺度气候因子对小流域影响不显著，未能提高模型预报能力。【结论】SAO-LSTM模型预报能力显著提高，具有优秀的泛化能力和鲁棒性；对于小流域的径流预报，降雨是关键驱动因素，重要性显著优于其他特征变量，而大尺度气候因子贡献较弱。研究提出的SAO-LSTM模型具有更高的预报精度，为理解影响径流的关键因素提供思路，在泛化能力方面表现出色，具有很好的应用场景，可为防洪和干旱决策提供一定的模型支撑。

Abstract

[Objective] Due to the frequent occurrence of disasters triggered by extreme runoff events under the influence of climate change and human activities, Long Short-Term Memory(LSTM)—a deep learning model—has been widely applied for runoff forecasting. However, it still requires improvements in both accuracy and interpretability. [Methods] A hybrid model for runoff forecasting was developed by combining the Snow Ablation Optimizer(SAO) with the LSTM model. Hydrometeorological characteristics of the watershed(runoff, precipitation, temperature) and six large-scale climate factors selected based on Pearson Correlation Coefficient(PCC) were used as model inputs. The model was compared with the Back Propagation(BP) neural network, and the Shapley value method was applied to analyze the importance and contribution of the input features. [Results] The SAO-LSTM model achieved a Mean Absolute Percentage Error(MAPE) of 0.26, a Coefficient of Determination(R²) of 0.80, and a Nash-Sutcliffe Efficiency(NSE) of 0.80, significantly outperforming both the LSTM and BP models and demonstrating excellent generalization ability. Shapley interpretation result indicated that precipitation was the key driving factor, while large-scale climate factors had no significant impact on small watersheds and failed to improve the model's forecasting performance. [Conclusion] The SAO-LSTM model significantly improves forecasting performance and exhibits excellent generalization ability and robustness. For runoff forecasting in small watersheds, precipitation is the key driving factor, with significantly higher importance than other feature variables, while large-scale climate factors contribute relatively little. The proposed SAO-LSTM model offers higher forecasting accuracy, provides insights into key factors influencing runoff, demonstrates excellent generalization ability, and shows promising application potential, thereby offering model support for flood control and drought decision-making.