A method of data analysis of temproal sequences,based on the replacement of new-old information of restrictive memory with lapse of time is developed. This appraoch is based on the modelling of the principal component analysis for the prior L square matrix of period extrapolation matrix Of mean generating function only,with a few principal components inclouding main information of temporal sequence.The scheme is applied to the multi-step forecast of annual meantemperatures of 28 cities in Sichuan,and the relative errors between forecast and real values are not more than 14%,and it is shown that the scheme is effective for the multi-step forecast of temperature.