At present, the forecasting model of temperature is mainly dependent on the number of monitored data, the prediction effect is not good as the monitored data are less. The paper attempts to set up and improve variable dimension fractal of forecasting model based on fractal theory, and analyzes the monitored data of Bailianya Arch Dam ' s temperature. The forecasted result shows that the improved model takes advantage of the feature of fractal theory self-similarity, and decreases the dependence on the data quantity and the noise influence on prediction effect. The model is of high precision, and has vast noise immunity.