Finding periodical regularities in sequential databases is an important topic in Knowledge Discovery. In pattern mining such regularity is modeled as partially periodic patterns, where typical periods (e.g., daily or weekly) can be considered. Although efficient algorithms have been studied, applying them to real databases is still challenging because they are noisy and most transactions are not extremely frequent in practice. They cause a combinatorial explosion of patterns and the difficulty of tuning a threshold parameter. To overcome these issues we investigate a pre-processing method called skeletonization, which was recently introduced for finding sequential patterns. It tries to find clusters of symbols in patterns, aiming at shrinking the space of all possible patterns in order to avoid the combinatorial explosion and to provide comprehensive patterns. The key idea is to compute similarities within symbols in patterns from a given database based on the definition of patterns we would like to mine, and to use clustering methods based on the similarities computed. Although the original method cannot allow for periods, we generalize it by using the periodicity. We give experimental results using both synthetic and real datasets, and compare results of mining with and without the skeletonization, to see that our method helps us to obtain comprehensive partially periodic patterns.