A hydrologic similarity-based parameters dynamic matching framework: Application to enhance the real-time flood...

Area of Science:

• Hydrology and Water Resources Engineering
• Environmental Science
• Data Science and Machine Learning

Background:

• Conventional hydrological models assume static, time-invariant parameters, which is increasingly questionable due to climate change and human interventions.
• Machine learning (ML) offers powerful tools for analyzing extensive hydrometeorological data to identify complex patterns.
• There is a need for advanced flood forecasting methods that can adapt to nonstationary hydrological conditions.

Purpose of the Study:

• To propose and evaluate a hybrid framework (HSPDM) for enhancing the precision of real-time flood forecasting.
• To dynamically retrieve time-variant hydrological model parameters using ML techniques for improved forecasting accuracy.
• To assess the framework's performance against traditional time-invariant parameter models and probabilistic forecasting schemes.

Main Methods:

• Developed a hybrid framework (HSPDM) integrating ML techniques like k-means, K-Nearest Neighbor (KNN), and embedding-based subsequence matching (EBSM).
• Dynamically extracted hourly time-variant hydrological model parameters from historical similar flood events.
• Compared three forecasting schemes: traditional time-invariant parameters, hourly time-variant parameters, and probabilistic forecasting.

Main Results:

• The HSPDM framework successfully identified continuous flood subsequences with high accuracy and acceptable computational time.
• The hourly time-variant parameter scheme demonstrated superior forecasting accuracy, outperforming the traditional time-invariant parameter scheme.
• Forecasts from the time-variant scheme aligned with and even surpassed probabilistic forecasts, indicating enhanced reliability.

Conclusions:

• The proposed hybrid framework advances real-time flood forecasting by incorporating hourly time-variant parameters, improving understanding of model behavior under nonstationary conditions.
• This approach offers a robust alternative for operational flood control and disaster mitigation.
• The study highlights the potential of ML-driven dynamic parameterization in hydrological modeling.