Research on machine learning hybrid framework by coupling grid-based runoff generation model and runoff process...

Area of Science:

• Hydrology
• Water Resource Management
• Computational Science

Background:

• Machine learning models like Long Short-Term Memory (LSTM) are crucial for watershed flood forecasting and disaster reduction.
• Traditional LSTM models often underestimate peak flows and lack robustness in flood prediction applications.

Purpose of the Study:

• To develop an improved flood forecasting framework addressing LSTM limitations.
• To integrate a Grid-based Runoff Generation Model (GRGM) and a runoff process vectorization (RPV) method with LSTM.
• To enhance the accuracy and robustness of flood prediction models.

Main Methods:

• A novel hybrid deep learning framework, GRGM-RPV-LSTM, was developed by coupling GRGM, RPV, and LSTM.
• The GRGM model was used to simulate runoff considering spatial distribution.
• The RPV method was employed to capture time series characteristics of runoff processes (rising, peak, recession).

Main Results:

• The GRGM model demonstrated high accuracy in runoff simulation with a relative error of 8.41% and a coefficient of determination of 0.976.
• The GRGM-RPV-LSTM model achieved a Nash efficiency coefficient greater than 0.9, outperforming LSTM and GRGM-LSTM models.
• The hybrid model showed superior peak flow prediction accuracy and robustness, especially with increasing lead times.

Conclusions:

• The GRGM-RPV-LSTM framework significantly improves flood forecasting accuracy and robustness compared to existing models.
• Considering spatial runoff patterns and time series characteristics enhances prediction capabilities.
• This research provides a scientific basis for effective flood control and disaster reduction in watersheds.