Abstract: It is difficult to achieve trend forecasting by the traditional prediction theory, because the state parameters of hydropower generating units are nonlinear, non-stationary, and with a small sample size available. Therefore, a time series combination model has been developed in this study. Wavelet transform can focus into any details of the signal and decompose a state sequence into a non-linear trend part and stationary fluctuating parts. By applying such decomposition to the vibration state sequences of hydropower generating units, a Least Square Support Vector Machine (LSSVM) prediction model was used for the trend part, and an Auto Regressive (AR) model for the fluctuating parts. The forecasting outcomes of these models were integrated by the principle of superimposition to achieve a final prediction. A case study of vibration state sequences shows that the forecasted and measured values agree well. Thus, the combination model presented in this paper achieves a high accuracy and the results would be useful for motivating the status early-warning for hydropower generating units.