Informed Neural Networks for Flood Forecasting With Limited Amount of Training Data

Abstract

This study introduces a novel method called Informed Neural Networks (INNs), developed to enhance flood forecasting accuracy, particularly under limited data conditions. Accurate flood forecasts are crucial for timely evacuations, especially as heavy rainfall increasingly threatens areas previously unaffected by flooding. Traditional methods often require extensive data and frequent updates, making them costly and challenging to maintain. INNs address these challenges by enabling accurate predictions under limited data conditions. We propose an INN architecture for rivers in regions like Japan, where floods are predominantly caused by rainfall. We applied the INN to both rainfall-dominated and non-rainfall-dominated floods to evaluate its effectiveness and limitations. Our experiments show that the INN effectively integrates domain knowledge, maintains performance, and achieves lower prediction errors than ANN in data-scarce scenarios. These findings highlight the potential of INNs as a promising approach for future flood forecasting, particularly in data-limited environments.