Model identification of medium and longterm hydrologic forecast is studied in terms of pretreatment

data length and ways of modeling which are taken as primary factors for the prediction results. Based on finite sampling information criterion (FSIC)

combined information criterion (CIC) is utilized to choose the proper order of the model. Kalman filtering is also used for nonlinear prediction. It is concluded that: 1) In model identification

reasonability of the pretreatment should be tested through the prediction results from the model if it significantly reduces the complexity of the model. 2) Data length of modeling should be long enough to reflect inherent oscillations of the time series while excessive amount brings in extra complexity

more time-consuming and less robustness. 3) Sliding model is better for larger flux and the streamflow peaks prediction

and sacrifices the precise of predicting relatively low run-off. 4) Kalman filtering used as a prediction method of runoff can remarkably raise the forecast effects in any sections of the range with the accuracy rate of peakprediction up to 63.64%.