High-flow forecasting with deep learning and dynamic multimodal feature fusion

Accurate streamflow prediction is essential for effective water resource management, yet it remains challenging due to unexpected environmental conditions, rapid flow variability, and sensor limitations. This study explores novel fusion techniques for deep learning-based flow forecasting models, aiming to adapt fusion operations based on the data’s characteristics. Using two neural network baseline architectures we explored dynamic operation-level fusion and attention-based fusion to integrate heterogeneous and multisource data for univariate (single gauge) and multivariate (multi-gauge) forecasting. The study area is the Severn Basin in the UK, known for long medium- to high-flow periods and shorter low-flow intervals. Results show that dynamic operation-level fusion consistently improved on attention-based fusion for both univariate and multivariate models by a 3.96% lower MAE as well as a 7.40% lower MAE on the extremes. Also, multivariate models improved the MAE as well as the MAE for the extremes by 3% and 3.59% respectively, while achieving a significant reduction in training and inference times by 74%. However, univariate models showed twice as much lower error rates and greater stability. Furthermore, notably in two river stations, all models underperformed due to rapid flow variability and flashy hydrological responses in smaller catchment areas, suggesting in the future the use of higher-resolution climatic data. Overall, the study shows the potential of new dynamic multimodal fusion techniques, navigating the operational trade-offs between speed, stability, and accuracy across multi- and uni-variate training strategies in streamflow forecasting. Nonetheless, the need for an optimal operational balance remains, suggesting further refinement of fusion techniques and focusing on minimising uncertainty especially for the multivariate models.