Data-driven model as a post-process for daily streamflow prediction in ungauged basins

Area of Science:

• Hydrology
• Water Resource Management
• Environmental Science

Background:

• Streamflow prediction in ungauged basins (PUB) presents significant challenges for water resource planning.
• Existing data-driven models (DDMs) show promise but require further refinement for improved accuracy and applicability in PUB.
• Process-based models (PBMs) also face limitations in ungauged basin predictions.

Purpose of the Study:

• To propose and investigate a novel framework for enhancing PUB performance.
• To utilize data-driven models (DDMs) as post-processors for both PBMs and other DDMs.
• To assess the effectiveness of post-processing in improving streamflow prediction accuracy in ungauged basins.

Main Methods:

• Employed the Parsimonious EcoHydrologic Model (PEHM) as a PBM.
• Utilized Long Short-Term Memory (LSTM) and Random Forest (RF) as DDMs.
• Applied RF and LSTM as post-processors to refine streamflow predictions from PEHM and standalone DDMs across 28 Korean basins assumed to be ungauged.

Main Results:

• Streamflow predictions from PEHM and LSTM were initially generated for ungauged basins.
• Post-processing with RF significantly improved the accuracy of streamflow predictions.
• Evaluation of LSTM as a post-processor also indicated potential benefits for PUB.

Conclusions:

• The proposed framework demonstrates the significant value of DDM post-processing for enhancing streamflow prediction in ungauged basins.
• This approach offers a viable strategy to improve water resource management and planning in data-scarce regions.
• Further research into various post-processing techniques can lead to more robust and accurate PUB.