Comparing an Innovative Neural Networks-Based Method with Existing Methods for Streamflow Naturalization
Abstract
Natural streamflow data is required in many hydrological applications, such as frequency analysis and model calibration. This data is sometimes unavailable, either because of a lack of gauging stations, or because of human constructions like dams and reservoirs. There exist several streamflow naturalization methods that can be used to estimate natural streamflow series, but so far, none of them has consensually been recognized as universally efficient. Consequently, further research on the topic is needed. Here we hypothesize that artificial neural networks (ANN) could provide a good means to model the relationship between the physiographical characteristics of a basin and the parameters of a hydrological model. To verify this hypothesis, ANN were trained to represent this relationship using data from gauged basins, and once learned, allowed for the estimation of the hydrological model parameters for ungauged basins using only their physiographical characteristics. To account for uncertainty an ensemble of ANN were trained to estimate the parameters of the hydrological model GR4J coupled with the snow accumulation and melt routine CemaNeige. The proposed method was applied on a combined dataset of 671 basins in the USA from the CAMELS dataset and on 85 basins from Quebec, Canada. A k-fold cross-validation was applied to assess the robustness of the method. The results of the proposed method, in terms of reconstructing the natural streamflow, were also compared with two well-known streamflow naturalization methods: the Drainage Area Ratio (DAR) and spatial proximity transfer. The results indicate that the proposed method outperforms the DAR and spatial proximity transfer method, with residuals of estimated vs. calibration parameters having a median close to zero. Furthermore, simulated streamflow series obtained using the parameters obtained from ANN in the GR4J model have a lower (better) Continuous Ranked Probability Score (CRPS) than the mean absolute error (MAE) for the two other deterministic methods. It can be concluded that the proposed method is a promising new avenue to estimate natural streamflows in ungauged or regulated basins.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35ZB.02M