Annual precipitation forecast is necessary for drought early warning and taking effective measures to defend drought happen, but how to get more accurate predictive information is still the hotspot. As we known, precipitation often contains randomness interferential with the change of time, while cloud model could realize the transformation between uncertain concept and quantitative values, which is built up on the basis of traditional fuzzy mathematics and probability statistics, thus the volume of precipitation in the future can be achieved through mining association rules from historical and current data. From this, this paper put forward an improved cloud reasoning model to predict annual precipitation which is based on the pretreatment of wavelet de-noise technique and revise of theoretical frequency curve. After that, the predictive results can reflect inter-annual variation of precipitation better in Xi'an, and its accuracy has been greatly improved, the error can be maintained basically less than 30%.