Global increases in the occurrence of harmful algal blooms (HABs) are of major concern in water quality and resource management. A predictive model capable of quantifying the spatiotemporal associations between HABs and their influencing factors is required for effective preventive management. In this study, a feature stream network (FSN) model is proposed to provide daily forecasts of cyanobacteria abundance at multiple monitoring sites simultaneously in a river network. The spatial connectivity between monitoring sites was expressed as a directed acyclic graph comprising edges and nodes representing flows and monitoring sites, respectively. Furthermore, a segment-wise node connection structure was developed to extract the latent features of a river segment comprising individual nodes and sequentially transfer them to the downstream segment(s). In addition, a feature engineering–attention hybrid mechanism was employed to address temporal mismatches among different monitoring schemes while adding explainability to the model. Consequently, the FSN showed improved predictive performance, temporal resolution, and explainability for multi-site forecasts of HAB in a single model framework. The developed model was applied to a bloom-prone middle course of the Nakdong River, South Korea. Various hydrological, environmental, and biological factors were utilized for forecasting the cyanobacteria abundance. The FSN exhibited a high degree of accuracy across the sites for the test data with a coefficient of determination in the range of 0.64–0.71 and root mean square error in the range of 2.06–2.26 cells/mL on natural log scales. Although the relative importance of input features varied across the sites, the features extracted from nearby nodes consistently exhibited high importance in forecasting the cyanobacteria abundance. These explanations indicate that the proposed model can successfully characterize the spatial hierarchy of a river network. A scenario analysis suggested that reduced total nitrogen loads in the effluents from the wastewater treatment plant and the combined operations of upstream and downstream weirs were effective in managing HABs.

中文翻译：

---

开发基于深度学习的特征流网络，从网络角度预测河流有害藻华


有害藻华 （HAB） 的全球发生率增加是水质和资源管理的主要问题。有效的预防性管理需要一个能够量化 HAB 与其影响因素之间时空关联的预测模型。本研究提出了一种特征河流网络 （FSN） 模型，以同时提供河流网络中多个监测点的蓝藻丰度的每日预测。监测站点之间的空间连接表示为有向无环图，该图由分别表示流量和监测站点的边和节点组成。此外，开发了一种分段节点连接结构，以提取由单个节点组成的河段的潜在特征，并依次将它们传输到下游段。此外，采用了特征工程-注意力混合机制来解决不同监控方案之间的时间不匹配，同时增加了模型的可解释性。因此，FSN 在单个模型框架中显示出改进的 HAB 多站点预测的预测性能、时间分辨率和可解释性。开发的模型应用于韩国洛东江容易发生水华的中游。利用各种水文、环境和生物因素来预测蓝藻的丰度。FSN 在测试数据的各个站点上表现出高度的准确性，在自然对数尺度上，决定系数在 0.64-0.71 范围内，均方根误差在 2.06-2.26 个细胞/mL 范围内。 尽管输入特征的相对重要性因地点而异，但从附近节点提取的特征在预测蓝藻丰度方面始终表现出高度重要性。这些解释表明，所提出的模型可以成功地描述河流网络的空间层次结构。情景分析表明，减少污水处理厂出水中的总氮负荷以及上游和下游堰的联合操作可有效管理 HAB。