Abstract: Climate systems are generally non-stationary. The data processing of climate observations, e.g. multi-step prediction of temperature and precipitation, is usually performed under an assumption that the time series is stationary. This may lead to unsatisfied accuracy in the prediction. In this paper, we study on a new method integrating an ensemble prediction technique and a stepwise regression model based on Mean-Valued Generated Function. By using the Empirical Mode Decomposition (EMD), we decomposed a non-stationary time series into a series of Intrinsic Mode Function (IMF) components and a trend component which is preliminary stable. Then a prediction model was built for each IMF component using an ensemble prediction technique and a stepwise regression analysis. The final results can be obtained by linear fitting. Using this method and the Visual Studio 2008, we developed a short-term climate prediction system which is calibrated by the temperature anomaly from 88 meteorological observation stations in Guangxi Zhuang Autonomous Region during the Februarys of 1957—2005. Compared to the normal and single prediction, the new method involving the EMD and the ensemble prediction technique has better performance in the prediction of climate change trend and abrupt climate changes when the multi-step prediction is made using historical data.