Abstract: In this paper，a new time-serial prediction model called relaxing wavelet neural-net is analyzed.At the same time，it decides the length of sample serial by using a kind of statistic method called variance analysis so that the knot number of input layer can be efficiently decided，which is based on analyzing time-serial prediction by means of neural-net.Using this model，it operates the annual average sunspot serial through wavelet-decomposing，reconstructing，prediction and composing.Then，the total prediction effect is gotten.At the same time，the prediction effect about new model and traditional BP-net is compared and the substaintial cause of prediction difference of both types is analyzed.In total，this paper reflects a kind of thought which simplifies complex problem and combines wavelet’s multi-discrimination with neural-net’s non-linear approaching function.This idea exerted wavelet-transformation and neural-net’s advantage respectively.Thus，the prediction precision is obviously enhanced.