FlowNet: A Generic Independent and Interactive Model for Streamflow Forecasting

ICLR 2026 Conference Submission17397 Authors

Keywords: time series forecasting, deep learning, machine learning

TL;DR: FlowNet: A Generic Independent and Interactive Model for Streamflow Forecasting

Abstract: Streamflow forecasting plays a crucial role in water research for flood prevention, water resource management, or climate resilience. However, it is a challenging task due to complex hydrological system interactions, human interventions and global climate change. In this paper, we introduce FlowNet, a \emph{unique local global interactive modeling} framework, which is capable of effectively predicting multiple hydrology stations with varied input climate features and data availability at the same time. The key idea of FlowNet is to contruct \emph{independent} prediction models for each station from its local data and from its adjacent neighbors via a hydrological-related directed graph before letting these models to \emph{iteratively} and \emph{interactively} adjust each other to maximize their prediction agreements. This helps to reduce uncertainty, thus improving their accuracy. Additionally, FlowNet dynamically captures inter-station relationships via its directional and delay-aware graph reconstruction method. As a generic framework, FlowNet can be used with any existing Deep Learning (DL) backbone models such as RLinear, PatchTST or iTransformer. However, we also introduce another backbone, called Disentangled Multiscale Cross-attention Transformer (DMCT), to capture the multiscale seasonality-trend information for further performance boost. Extensive experiments on 3 large datasets, including LamaH (with 425 hydrology stations in Europe), CAMELS (672 stations in USA) and MRB (with 26 gauge stations in the Mekong River Basin), show that FlowNet significantly outperforms 18 state-of-the-art (SOTA) prediction methods in terms of MAE, RMSE, and NSE.

Primary Area: learning on time series and dynamical systems

Submission Number: 17397