Colambage et al. (2026) Integrating Machine Learning with Hydrological Modelling for Event-Based Streamflow Prediction: A Case Study of Bulimba Creek Catchment, South East Queensland

Identification

• Journal: Lecture notes in civil engineering
• Year: 2026
• Date: 2026-01-01
• Authors: Achini Colambage, Zhongzheng Wang, Buddhi Wijesiri, Jayaram Pudashine, Prasanna Egodawatta
• DOI: 10.1007/978-3-032-18708-6_29

Research Groups

• Queensland University of Technology, Brisbane, Australia
• Bureau of Meteorology, Melbourne, Australia
• Institute of Environment and Ecology, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China

Short Summary

This study develops a hybrid modelling framework integrating the conceptual Unified River Basin Simulator (URBS) with a Long Short-Term Memory (LSTM) neural network as a residual post-processor to improve event-based streamflow prediction. The framework significantly enhances hydrograph accuracy and provides reliable uncertainty estimates, demonstrating its potential for operational flood forecasting in Australian catchments.

Objective

• To introduce and evaluate an event-based hybrid modelling framework that integrates the Unified River Basin Simulator (URBS) with a Long Short-Term Memory (LSTM) neural network, implemented as a residual post-processor, to predict and correct systematic errors in URBS-simulated discharge while retaining the model’s physical foundation.

Study Configuration

• Spatial Scale: Bulimba Creek catchment, Brisbane, South East Queensland, Australia
• Temporal Scale: Event-based, using 15-minute rainfall and discharge data; ten historical storm events for cross-validation and one multi-peak event for independent testing.

Methodology and Data

• Models used: Unified River Basin Simulator (URBS) (conceptual hydrological model), Long Short-Term Memory (LSTM) neural network (residual post-processor), Monte Carlo dropout (for uncertainty quantification).
• Data sources: Observed 15-minute rainfall and discharge data for the Bulimba Creek catchment.

Main Results

• The hybrid model significantly improved performance across development events, increasing Nash-Sutcliffe Efficiency (NSE) from 0.34–0.87 (URBS) to 0.79–0.97.
• Kling–Gupta Efficiency (KGE) improved by 0.20–0.40, and Root Mean Squared Error (RMSE) was reduced by 40–70%.
• For an independent test event, NSE increased from 0.88 to 0.97, and volume bias improved from -28% to -5%.
• The framework successfully produced calibrated Monte Carlo dropout uncertainty bands, providing reliable confidence intervals for flood forecasts.

Contributions

• Demonstrates, for the first time in an event-based, high-resolution setting, that hybrid residual learning can substantially enhance hydrograph accuracy and strengthen operational flood forecasting in Australian catchments.
• Introduces a robust method for integrating machine learning with conceptual hydrological models that maintains interpretability while significantly improving prediction accuracy and providing uncertainty estimates.

Funding

• Computational resources and services were provided by the eResearch Office, Queensland University of Technology, Brisbane, Australia.

Citation

@article{Colambage2026Integrating,
  author = {Colambage, Achini and Wang, Zhongzheng and Wijesiri, Buddhi and Pudashine, Jayaram and Egodawatta, Prasanna},
  title = {Integrating Machine Learning with Hydrological Modelling for Event-Based Streamflow Prediction: A Case Study of Bulimba Creek Catchment, South East Queensland},
  journal = {Lecture notes in civil engineering},
  year = {2026},
  doi = {10.1007/978-3-032-18708-6_29},
  url = {https://doi.org/10.1007/978-3-032-18708-6_29}
}