The findings of this study demonstrate the potentialities in using the K-means algorithm for grouping data points related to cooling variables of LCP units. Additionally, the results show that it is suitable to divide the data points into four clusters. The identified clusters differ with regards to variables, among other, such as LCP return air temperature and temperature difference between chilled water supply and return. This is beneficial in identifying undesired operational statuses of LCPs, e.g., low temperature difference between chilled water supply and return, which is an indicator of a poor cooling performance. Clusters 1 is characterized by a combination of low LCP return air temperature and low average cooling power, which can be attributed to nonoperational periods during large parts of the analyzed time period. Cluster 2 has moderate LCP return air temperature, relatively low chilled water flow rate, and high △Tchilled water. In contrast, Cluster 3 demonstrates high chilled water flow rate and LCP return air temperatures, with relatively low △Tchilled water. Finally, Cluster 4 is featured by high LCP return air temperature, rather high △Tchilled water, and chilled water flow rate. It should be highlighted that in the context of energy efficiency, it is preferable to have a high △Tchilled water, and a low chilled water flow rate, meaning that Cluster 4 is preferred compared to Cluster 3.

With regards to the use of K-means as method in this research, it enhances data visualization and aids in deeper understanding of complex patterns within a dataset. Consequently, K-means can be used as a tool for data-driven analysis of cooling patterns in ICT rooms. Within the context of this research project, the use of K-means has been key for communication of results to facility management consultants and employees at Ericsson AB. Hence, undesired cooling patterns that deviate from the desired ones can be effectively communicated. Moreover, it is important to address the selection of four clusters, instead of three clusters, which were also considered suitable as previously mentioned. The motivation for this is to obtain a more detailed and comprehensive representation of the cooling characteristics in the dataset. As a result, this allows for a more granular depiction of the cooling patterns in the investigated dataset.