Abstract: The ensemble Kalman filter approach has been used to correct the state variable in hydrological models. Difficulties of its application include how to select the state variable for correction, whether or not to synchronize parameter correction with the state variable, and how to set up the filter algorithm's hyperparameters. To address these issues, we take the calibrated GR5J model for the Qijiang River basin as a prototype tool to assimilate observed streamflows and correct model state variables using feedback correction. We use synthesis experiments and rolling forecast tests to examine the impacts of state variable selection, model parameter disruption, and hyperparameter optimization of the filter algorithm on forecast accuracy. The results suggest that while the biased initial state could be specified, the ensemble Kalman filter does raise forecast accuracy; otherwise, a better way is to fix the runoff generation variable and the flow confluence variable simultaneously to avoid overcorrection on model states. In the case of biased model parameters, it is best to identify the parameter first and then adjust the state variable. Increasing the ensemble members and warm-up periods generally improve correction accuracy, but the impacts of model noises and observation noises on the correction accuracy are non-monotonic. The filter algorithm is superior to the warm-up method, though its forecast accuracy decreases with an increasing forecast period. The findings would help apply the state correction method to operational forecasting.

Key words: ensemble Kalman filter, hydrological model, flood forecast, data assimilation, real-time correction