This paper discusses the construction and application of BP-type multilevel mapping model with dynamic learning rateηBased on the 2002 observed temperature data of Yuanjiang Meterorological Station in Hunan Province,daily mean temperature prediction is made by using a dynamic learning rate model,a fixed learning rate model and a stepwise regression model.Results suggest that the BP neural network model is superior to the regression model in predictions,and also swift in learning if the dynamic learning rate is adopted.