| 摘要: |
| 对湖泊总磷的变化预测和来源识别对水资源调度和流域生态治理有着重要的意义,然而复杂的生物和化学反应导致的总磷(TP)浓度变化的非稳态特性给总磷的准确预测带来极大的困难。为克服这一挑战,本文引入了Seasonal and Trend decomposition using Loess(STL)技术和SHapley Additive exPlanations(SHAP)结合长短期记忆网络(LSTM)和门控制单元(GRU)构建了可解释框架STL-LSTM-SHAP(SLSEF),以增强神经网络模型的预测能力并且为预测结果提供合理可靠的解释。在中国骆马湖(Lake Luoma),该框架总磷预报精度R2为0.884,相比较单一LSTM和CNN-LSTM提高了5%和15.7%,当预测时间步长增加到8h时,SLSEF对总磷的趋势预报仍能保持较高的预测精度(R2>0.8)。与以往研究不同的是,本研究不仅预测湖泊未来总磷的演变趋势,还通过计算不同特征值对预测结果的贡献权重,结合MIKE模型,验证了总磷浓度与不同特征值之间的时空响应规律。解释结果表明运河来水是骆马湖总磷浓度最重要的影响因素,并且不同断面的污染源受水动力等因素的影响存在显著的时空差异。本文凸显了神经网络模型在预警水体污染方面的可实施性,并且为提高传统神经网络的学习能力和可解释性的开发与验证提供了重要方向。 |
| 关键词: 深度学习 LSTM SHAP 结构方程 总磷 可解释性研究 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金（41971137、U2240224、42001109） |
|
| Research on Interpretable LSTM for predicting total phosphorus in Lakes |
|
dingyiding1, Fan Hongxiang2, Xu Ligang, Jiang Mingliang, Lv Haishen, Zhu Yonghua, Chen Junxiang
|
|
1.College of Hydrology and Water Resources, Hohai University;2.Nanjing Institute of Geography & Limnology Chinese Academy of Sciences
|
| Abstract: |
| The prediction and source identification of total phosphorus (TP) in lakes have significant implications for water resource management and watershed ecological governance. However, the non-stationary characteristics of TP concentration changes, driven by complex biological and chemical reactions, pose substantial challenges for accurate prediction. To address this challenge, this paper introduces the Seasonal and Trend decomposition using Loess (STL) technique and SHapley Additive exPlanations (SHAP), combined with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), to construct an interpretable framework known as STL-LSTM-SHAP (SLSEF). This framework enhances the predictive capabilities of neural network models and provides reasonable and reliable explanations for the prediction results.In the case of Lake Luoma in China, the SLSEF framework achieves a total phosphorus forecasting accuracy with an R-squared (R2) value of 0.884, which is a 5% and 15.7% improvement compared to a single LSTM and CNN-LSTM. Even when the prediction time step increases to 8 hours, SLSEF maintains a high prediction accuracy for total phosphorus trends (R2 > 0.8). Unlike previous research, this study not only predicts the future evolution trends of total phosphorus in the lake but also validates the spatiotemporal response patterns between total phosphorus concentration and various feature contributions using the MIKE model. The interpretive results indicate that canal inflow is the most critical factor influencing Lake Luoma"s total phosphorus concentration, and there are significant spatiotemporal variations in the pollution sources at different sections, influenced by hydrodynamic factors and other elements. This research underscores the feasibility of neural network models in early warning of water body pollution and provides an important direction for the development and validation of improved learning and interpretability in traditional neural networks. |
| Key words: deep learning LSTM SHAP SEM Total phosphorus Interpretability |
