Deep learning for Australian streamflow predictions
Project leader: Stephanie Clark
Project description
This project aims to enhance our understanding of the contexts in which deep learning has the potential to add value within Australian hydrological modelling—particularly in improving predictions in large or data-sparse catchments, enhancing generalisation to changing conditions, and supporting more efficient, scalable modelling approaches.
Deep learning methods provide adaptable and accurate models that promise valuable new pathways for gaining insight into hydrological processes. In this study, practical trials of deep learning time series models – Long short-term memory networks (LSTMs) – are conducted, to evaluate their performance in streamflow prediction under Australia’s dry and variable conditions. Covering a wide range of catchments, the study assesses the LSTMs’ prediction accuracy and scalability in diverse climate and landscape environments. The project is conducted in two stages:
- Basin scale – the prediction performance of deep learning models is compared to traditional models in almost 500 Australian basins, demonstrating the ability of deep learning models to efficiently meet or exceed traditional modelling performance in the majority of catchments.
- Continental scale – a single continent-wide deep learning model is developed to predict streamflow simultaneously at hundreds of basins across Australia, demonstrating the scalability benefits. Patterns of deep learning model performance are investigated for different Australian catchment types.
Project impact
The results highlight the potential of deep learning methods to advance water resource management across the diverse hydrologic conditions in Australia by enabling faster model deployment at scale, offering new insights through the use of nationally and internationally consistent datasets, and improving prediction accuracy in data-sparse regions and under challenging conditions such as climate change. This provides strong support for the integration of deep learning models into the national hydrological toolkit, demonstrating their effectiveness in delivering efficient and accurate streamflow modelling.
Project publications
Associations between deep learning runoff predictions and hydrogeological conditions in Australia
Contact
If you would like to know more about this project, please contact Stephanie Clark (https://people.csiro.au/c/s/stephanie-clark)
