Role of multimodel combination and data assimilation in improving streamflow prediction over multiple time scales

Abstract

In hydrologic modeling, various uncertainty sources may arise due to simplification/representation of real-world spatially distributed processes into the modeling framework, such as uncertainty due to model structure, initial conditions and input errors. One approach that is currently gaining attention to reduce model uncertainty is by optimally combining multiple models. The rationale behind this approach is that optimal weights could be derived for each model during the model combination process so that the developed multimodel predictions will result in improved predictability. Another approach—data assimilation—is gaining popularity in reducing uncertainty by deriving updated initial conditions recursively from the current available observations to reduce overall uncertainty by minimizing the error covariance matrix of state variables. In this paper, an experimental design is proposed to test the performance of both approaches, multimodel combination and data assimilation, in improving the hydrologic prediction at daily and monthly time scales. The experimental design is constructed on a synthetic basis such that the ‘true’ model structure and streamflow values are known. We evaluated the performance of multimodel combination and data assimilation through the experimental design at monthly and daily time scales, then compare how uncertainty due to initial conditions and hydrologic model can be dominant at the respective time scales. For the multimodel combination, we combined the models by evaluating the model performance conditioned on the predictor state. For data assimilation, the Ensemble Kalman Filter (EnKF) was adopted to test its usefulness through the same experimental design. Results from the synthetic study showed that under increased model uncertainty, the multimodel algorithm consistently performed better than the single model predictions and the EnKF algorithm in terms of all performance measures at monthly time scale. However, under daily time scale, the multimodel algorithm did not performing better than the EnKF algorithm in most of the model uncertainty cases. Findings from the synthetic study was also consistent upon application in predicting streamflow at daily and monthly time scales for a watershed in North Carolina.