A river network model using a weight-based merged LSTM for multi-source monitoring integration

Rivers typically exhibit spatial connectivity from upstream to downstream, with various heterogeneous monitoring systems operating concurrently across basins. While graph neural networks (GNNs) have shown promise in modeling spatial connectivity, they remain limited by reliance on features common to all nodes. As the diversity of monitoring systems increases, the overlap of measured variables decreases, reducing usable input features and limiting model applicability. To address this issue, this study proposes a river network model based on a weight-based merged long short-term memory (LSTM) architecture for forecasting daily total organic carbon (TOC) concentrations by integrating multi-source data from spatially connected monitoring sites. The river network is divided into upstream, midstream, and downstream segments, with each input processed independently and merged into a unified representation. Segment models are connected in sequence to represent directional flow. Trainable scalar weights are included to quantify the relative contribution of each site and enhance spatial interpretability. The model is applied to a section of the Han River in South Korea, which flows through the Seoul metropolitan area in South Korea. The model demonstrates strong forecasting performance, with mean absolute errors ranging from 0.055 to 0.518, root mean squared errors from 0.075 to 0.784, and coefficients of determination between 0.424 and 0.721. Scenario analyses using site-specific contributions are conducted to evaluate changes in TOC concentrations under pollution reduction scenarios. This river network modeling framework is adaptable to a wide range of applications and provides practical utility for watershed-scale water quality forecasting and management. • A river network model was proposed for TOC prediction in a connected river system. • The model adopted a weighted merged LSTM structure to integrate multi-source data. • The model achieved a maximum R 2 of 0.721 for TOC prediction. • Feature-flexible design effectively accommodated heterogeneous inputs across sites. • Site-specific pollution contributions were quantitatively assessed.