The Fukushima Daiichi accident demonstrated that risk assessment based solely on unit-by-unit treatment may be insufficient for nuclear power plant sites with multiple reactor units. At such sites, offsite consequence assessment may need to account for radioactive releases from more than one unit, making the treatment of multi-unit source terms an important issue in site-level analysis. A major difficulty in this context is the rapid increase in the number of possible source term category (STC) combinations as the number of units grows, which makes exhaustive evaluation computationally burdensome. Previous studies have addressed this issue using approaches such as binning based on logic tree and site level release category grouping, while categorization based on clustering using detailed source-term data has mainly been examined for single-unit applications. This paper presents a method for reducing multi-unit source term combinations using release characteristics and timing features. Unit-level source terms are represented using physically interpretable descriptors derived from severe accident analyses, including release magnitude, radionuclide composition, release timing, release duration, peak release characteristics, and release height. These descriptors are combined into multi-unit feature vectors, and unsupervised clustering based on k-medoids is applied to group similar combinations and select representative combinations for subsequent consequence calculations. The proposed method is illustrated through a case study based on the Hanul nuclear power plant site in South Korea, reflecting diverse reactor designs including FRAMATOME, OPR1000, and APR1400. By directly using release characteristics from MELCOR analyses, the proposed approach provides a practical and interpretable framework for handling the complexity of multi-unit source term combinations in site-level consequence assessment.