Virtual Metrology (VM) is a method to conjecture manufacturing quality of a process tool based on data sensed from the process tool and without physical metrology operations. Historical data is used to produce the initial VM models, and then these models are applied to operate in a process drift/shift environment. The accuracy of VM highly depends on the modeling samples adopted during initial-creating and on-line-refreshing periods. Since design-of-experiments (DOE) may not be performed due to large resources required, how could we guarantee stability of the models and predictions when they move into these unknown environments? Conventionally, static-moving-window (SMW) schemes with a fixed window size are adopted during the on-line-refreshing period. The purpose of this paper is to propose a dynamic-moving-window (DMW) scheme for VM model refreshing. The DMW scheme adds a new sample into the model and applies a clustering technology to do similarity clustering. Next, the number of elements in each cluster is checked. If the largest number of elements is greater than the predefined threshold, then the oldest sample in the cluster with the largest population is deleted. Test results show that the DMW scheme has better on-line conjecture accuracy than that of the SMW scheme.