Abstract: To effectively utilize information of the section interdependence in the time series of annual runoff, a threshold auto-regressive (TAR) model is proposed to predict annual runoff.A simple and general scheme is presented to establish a TAR model.With an improved genetic algorithm, both the threshold values and auto-regressive coefficients can be optimized, and the problem of TAR modeling resolved, giving a powerful tool for wide application of the TAR model.A case study shows that the scheme is practical and efficient, and the TAR model can successfully reduce model errors and, by controlling the threshold values, ensure good stability and accuracy of the model forecast.As a general method, the scheme has theoretical value and wide range of applications in nonlinear time series prediction.

Key words: threshold auto-regressive model, prediction, annual runoff time series, genetic simulated annealing