Flood resilience through hybrid deep learning: Advanced forecasting for Taipei's urban drainage system

Area of Science:

• Environmental Science
• Hydrology
• Artificial Intelligence

Background:

• Climate change intensifies extreme rainfall, stressing urban drainage systems and increasing flood risks.
• Effective flood mitigation requires optimized operations and responsive disaster management.

Purpose of the Study:

• To develop a Real-Time Urban Drainage Early Warning System using deep learning for enhanced flood management.
• To integrate diverse data sources using knowledge graphs for a comprehensive view of flood dynamics.

Main Methods:

• Applied Convolutional Neural Networks combined with Back Propagation Neural Networks (CNN-BP) for forecasting.
• Developed multi-input multi-output multi-step (MIMOMS) models for sewer and water level predictions.
• Utilized knowledge graphs to integrate varied data sources for flood dynamics analysis.

Main Results:

• The CNN-BP model achieved high accuracy: R² of 0.97 for sewer water levels and R² of 0.99 for internal/external water levels at T+1.
• Demonstrated superior forecast accuracy in capturing water level trends during intense rainfall.
• Validated the model's capability for reliable real-time responsiveness.

Conclusions:

• The CNN-BP model significantly enhances urban flood prediction and early warning systems.
• Accurate forecasting enables optimized pump operations and intelligent flood control practices.
• The study provides a robust framework for effective environmental management in urban drainage systems.