AI improves long-term flood forecasting

While Australia may be known for its heatwaves, flooding is one of the country’s deadliest natural hazards, with devastating effects on communities and infrastructure. Floods also carry a significant financial cost, resulting in billions of dollars in insured losses and lost economic activity each year.

Combined with urban development, the worsening climate crisis is putting communities increasingly at risk as storms and flooding become more frequent and severe.

Addressing this issue was a key motivation for Professor Gandomi, an expert in machine learning and big data analytics, as better forecasting can help authorities shift from reactive to proactive flood management.

“If we can provide reliable information weeks rather than days ahead, authorities have more opportunity to prepare infrastructure, plan emergency responses and protect communities,” he said.

In the study, the team – including Marjan Kordani, Dr Mohammad Reza Nikoo, Mahmood Fooladi, Dr Iman Ahmadianfar and Professor Rouzbeh Nazari – used two decades of historical precipitation and streamflow data from nine precipitation stations and seven streamflow stations in the DuPage River basin in Illinois, USA. 

The researchers developed advanced deep learning models incorporating attention mechanisms to identify important patterns in rainfall and river flow data. They then combined the models’ predictions using a fusion approach to improve forecasting accuracy. 

The results demonstrated substantial improvements over the conventional machine learning approaches evaluated in the study. The ensemble deep learning models reduced prediction errors by approximately 38 per cent, while the proposed fusion framework achieved an approximately 47 per cent improvement in predictive performance, as reported in the study.