Error-correction of streamflow predictions from a global hydrological model using random forests

Abstract:

To improve streamflow predictions, researchers have implemented hybrid models that complement physically-based models with machine learning methods, in which the simulated streamflow and meteorological variables are used as predictors. Few studies however have included an extensive set of driving and state variables simulated by the model. Also, due to uncertainty in forward simulation models, errors in predictions are inevitable. We developed and evaluated a random forests (RF) based approach to correct predictions from the hydrological model PCR-GLOBWB (https://github.com/co822ee/PCR-GLOBWB_error-correction).

Meteorological input and an extensive list of state variables were used as predictors in the RF to estimate errors of streamflow predictions, which were then applied to correct simulated hydrographs. The RF was trained and applied separately at three gauging stations in the Rhine basin with different physiographic characteristics. The data from 1981-1990 was used for training, and the performance of the RF model was validated on 1991-2000.

The inclusion of state variables as predictors resulted in much better results compared to error correction using meteorological input only. Daily simulations from an uncalibrated PCR-GLOBWB (KGE 0.37-0.62 & NSE 0.19-0.39) were largely improved by applying the RF model (KGE 0.76-0.89 & NSE 0.64-0.80). The performances were equally good, when error-correction was applied on a calibrated PCR-GLOBWB run (KGE 0.72-0.87 & NSE 0.60-0.78). The performance with error-correction was better compared to the calibrated run without error-correction (KGE 0.57-0.71 & NSE 0.29-0.56). This indicates that the RF-based error-correction method is more efficient than calibration, particularly when the interest lies in minimizing streamflow discharge prediction error, such as in operational forecasting. The state variables that contribute most to error correction differ between catchments. The variables that are related to groundwater are informative in basins dominated by the aquifer sedimentary basins, while snow and surface water state variables are informative in snow dominated regions and near large lakes. We expect that the approach presented here can be extended to prediction at ungauged locations and to reconstruction of missing discharge data.