These years, modeling the process of Internet public opinions' evolution and trend forecasting based on that has become a hot topic. The existing short-term trend forecasting method ignores the variability of statistical properties of Internet public opinions’ evolution, which leads to a blind model selection, and the forecasting performance is poor. Therefore, this paper presents an adaptive evolution modeling method of Internet public opinions (AEMIPO). Firstly, this method tracks the statistical characteristics of the process of Internet public opinions' evolution dynamically, such as smoothness, periodicity and self-similarity. Then, by selecting ARMA, ARIMA, SARIMA and FARIMA, an alternative model bank is constructed. Finally, by making model selection rules, an appropriate model is selected to model and forecast the process of evolution adaptively. The experimental results show that compared with the existing methods, AEMIPO has higher forecasting accuracy and stability, and this method is more suitable for short-term modeling and trend forecasting of Internet public opinions’ evolution.