Case Study: Using eWater Source and Machine Learning to Forecast Australia’s Water Future

Two University of Sydney engineering students are examining the role of machine learning in forecasting dam behaviour and water supply resilience and how it can enhance the capabilities of established modelling tools such as eWater Source.

Two graduating civil engineering students at the University of Sydney, Zehua Zhang and Joash Yeap, have undertaken honours research projects aimed at improving the forecasting of dam behaviour and water supply resilience in an increasingly variable climate. Their projects explored the growing role of machine learning to enhance the capabilities of established modelling tools such as eWater Source.

Zehua Zhang investigated how machine learning methods could improve dam level forecasting across contrasting Australian catchments, while Joash Yeap undertook a detailed comparative analysis of conceptual hydrological models and machine learning approaches within a difficult to model catchment in northern New South Wales.

Machine learning models substantially outperformed traditional approaches, with the XGBoost model achieving particularly strong predictive performance under Australian conditions. Zehua also developed an automated forecasting framework capable of continual model retraining and updating, reducing the need for time-consuming manual calibration while improving operational forecasting capability. The work highlighted how machine learning can assist water authorities in anticipating storage declines, preparing for drought conditions, and strengthening long-term water supply resilience under climate uncertainty. By supporting research like this, eWater is seeking to understand how to leverage the benefits of machine learning with those of traditional modelling approaches, to provide the best possible tools for our community.

Figure 1 Detailed automated machine learning framework (Zhang, 2025)

Joash Yeap’s thesis examined how traditional conceptual hydrological models perform in difficult real-world catchments compared with machine learning approaches. His case study focused on the Rocky Creek Dam catchment, a system known for its difficult low-flow hydrology. Using eWater Source and its inbuilt GR4J rainfall-runoff model, Joash developed a conceptual hydrological model and directly compared it with an XGBoost machine learning model trained on the same hydro-climatic data.

The comparison revealed major differences between the modelling paradigms. The GR4J conceptual model performed poorly in the low-flow environment, producing inaccurate results under the chosen calibration conditions. In contrast, the XGBoost machine learning model demonstrated a stronger ability to identify and reproduce underlying hydro-climatic patterns within the catchment.

Joash’s work highlighted an important tension within modern hydrological modelling: while machine learning models may achieve superior numerical performance, they are often criticised as “black box” systems because they do not explicitly represent physical hydrological processes. This creates important questions for water authorities about trust, transparency and suitability for long-term climate adaptation planning.

This research demonstrates that the future of water forecasting may increasingly involve hybrid approaches that combine physically based models with data-driven machine learning systems. Traditional hydrological models offer interpretability, physical realism and policy transparency, while machine learning approaches offer adaptability, automation and improved predictive accuracy under highly variable conditions.

Both projects reinforce the value of advanced modelling literacy within civil and environmental engineering education. By working directly with eWater Source, GIS platforms, hydrological datasets and machine learning frameworks, the students developed skills directly relevant to the future management of Australia’s water resources.