Abstract:

Slow feature analysis (SFA) is an effective method to extract slowly varying information from known non-stationary time series.In this paper, the ability of the SFA method to extract slow feature information was first tested by using the Logistic non-stationary time series model, and then the slow feature signals extraction and prediction of monthly temperature anomalies series were carried out at Harbin station as a representative station in Heilongjiang province.The results showed that the SFA method can effectively extract the slow feature signals from the temperature anomalies series in Harbin.The extracted slow feature signals can reflect the trend, extreme value, and other information of the original temperature series.Fitting and prediction experiments show that the temperature prediction model after the introduction of SFA signals can improve the prediction ability, compared with the stationary model.The same conclusion is obtained for the independent sample predictions for the latest 48 months.

Key words: Slow feature analysis (SFA), Non-stationary time series, Temperature prediction model