Drifting Learning (DL) is an effective method to solve the regression problem in the field of data mining. The approach is established based on the combination of Local Weighted Learning (LWL) algorithm and Statistical Learning Theory (SLT). It is shown from the theoretic analysis and simulation that better performance on estimation precision and generalization ability than the traditional methods can be achieved. And this method is suitable for modeling complex industrial process with multiple work modes. In the algorithm, the optimized bandwidth selection is a key factor on the generalization performance and real-time performance. This paper first analyzes the effect of the optimized bandwidth on drifting learning method based on theoretic analysis and simulation, and then provides a novel optimized bandwidth selection algorithm. The simulation results show that the proposed approach can achieves performance superior to the existed methods.