The conceptual nature of rain-runoff models causes a significant drawback in the form of model bias. In certain types of applications, like real-time prediction, the model bias is directly mapped to the prediction result. The methodologies that are developed to improve real-time prediction, like data assimilation or real-time model calibration are taking the advantage of the information in the form of measured real-time data, but does not face the problem of model inaccuracy. In this paper the method based on data assimilation that is capable of detecting and reducing model bias (hence leading to more accurate real-time predictions) is proposed. The method has two principal steps. In the first step, the most recent model prediction is modified using observed data from the near past. The model prediction with reduced bias is obtained in this step. In the second step, the most recently measured runoff data value is assimilated with the model prediction from the previous step to further improve the prediction. This is done using the least-squares data assimilation technique. The new method is tested and verified on the case study of rain-runoff model of town Aarhus in Denmark. The results obtained show that the new method has improved prediction capabilities when compared to both off line model predictions and predictions of the model with data assimilation only. The additional benefit is the low computational cost since the new method requires only one rain-runoff model simulation run for a single prediction.
Branisavljevic, Nemanja; Hutton, Christopher; Kapelan, Zoran; Vamvakeridou-Lyroudia, Lydia; and Savić, Dragan A., "Real-Time Runoff Prediction Based On Data Assimilation And Model Bias Reduction" (2014). CUNY Academic Works.