In this paper, we propose a degree-of-freedom-based adaptive reduction method that ensures accuracy over a wide band. In the conventional dynamic condensation method, a single reduced model consisting of low-order modes is used throughout the analysis. This results in low accuracy in the high-frequency band because it does not reflect the characteristics of the frequencies. To address this issue, we implemented a reduced model for each frequency using a Taylor series. This method converts the transformation matrix into a frequency-independent form, which allows for a simple interpolation of the reduction model by updating the differences between frequencies. Numerical examples were adopted to examine the accuracy and efficiency of the proposed method.