A knowledge-based deep learning model for accurate urban drainage system prediction under spatiotemporally variable...

Area of Science:

• Environmental Engineering
• Hydrology
• Artificial Intelligence

Background:

• Accurate urban drainage water level prediction is crucial but challenged by sparse sensors and complex hydraulics.
• Traditional models require extensive data, while data-driven methods lack generalization.
• Existing methods struggle with observational sparsity and uncertain rainfall inputs.

Purpose of the Study:

• To develop a knowledge-based deep learning model for accurate and robust water level prediction in urban drainage systems.
• To integrate prior knowledge with sparse real-time monitoring data for improved forecasting.
• To enable real-time urban flood prediction and early warning systems.

Main Methods:

• A novel knowledge-based deep learning model was proposed, integrating physical insights with sparse monitoring data.
• The model propagates observational signals through network pathways for spatial and temporal correction.
• Evaluations were conducted using synthetic and real rainfall events under varying sensor coverage.

Main Results:

• The model achieved sub-decimeter accuracy across most nodes, maintaining high performance (NSE > 0.90) even with extreme sensor sparsity.
• Peak prediction errors were within 0.05 m in real-world applications, with high accuracy in overflow detection.
• Spatial residuals showed physical meaningfulness, aiding interpretation and sensor deployment strategies.

Conclusions:

• The knowledge-based deep learning model offers a scalable and robust solution for urban drainage management.
• It demonstrates superior stability and accuracy compared to purely data-driven approaches.
• The model's low input requirements and strong generalization ability support its deployment for real-time flood prediction and early warning.