Incorporating Rainfall Loss Process with Machine Learning Models for Prediction in Ungauged Basins
Abstract
Machine Learning (ML) and conceptual models are the two major strategies for streamflow prediction in ungauged basins (PUB). Researchers face the dilemma of selecting between ML and conceptual models for PUB due to the inevitable limitations in both approaches, such as the lack of interoperability of ML models and overparameterization of conceptual models. Theory-guided machine learning models can integrate ML models with conceptual models to complement the limitations of ML models. In this study, theory-guided ML models with theory-guided architecture (TGA) incorporate Random Forest models with the conceptual tutors: results of conceptual models and the soil parameters of Green-Ampt and SCS-CN models. Green-Ampt and SCS-CN models have poor performance in predicting direct runoff at field scale but can be improved by incorporating conceptual models with the residuals predicted by residual error models. TGA models with reconstructed Green-Ampt models and the soil parameters of Green-Ampt models significantly improve the ML models due to the successful adoption of infiltration processes with the dynamic soil information. In addition, the TGA models perform better in events with higher hydraulic conductivity. On the contrary, TGA models with SCS-CN-related conceptual tutors do not enhance the performance of ML models. Despite the higher accuracy of direct runoff predicted by SCS-CN models than the Green-Ampt-related models, the SCS-CN-related models fail to provide valuable information other than weather inputs for the TGA models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H35C..01L