Phase connections of end-users in electricity distribution are conventionally fixed during the first installation and remain so unless there is a specific problem or alteration in the network. However, a wrong phase connection may cause load imbalance, fluctuations, and loss in effi- ciency in energy distribution. Consequently, the advanced measurement and built-in advanced communication in smart meters create a strong potential for their application in continuously monitoring and analyzing electricity consumption patterns at the end-user level. In this work, machine-learning-voltage-based (MLV) techniques are used to identify phases in specific grid environments and to implement the method under real network voltage magnitude data from smart meters installed in the distribution network. The thesis will address this prob- lem by using Python as a tool to realize a way to explore the performance of MLV (specifically PCA and K-means) in identifying connections related to phase. Very few studies have been conducted that analyze the Principal Component Analysis (PCA)- based and constrained K-means clustering in chosen scenarios of the grid application. However, a research gap remains regarding the robustness and reliability of these algorithms in identi- fying phases correctly under varied grid conditions. More specifically, the confidence of such an algorithm to correctly identify phases across different grid settings has not been extensively investigated beyond PV-rich or constrained environments. This project will leverage resources from Sweden’s largest Distribution System Operator (DSO), and grid simulation experiments will be developed to generate datasets via smart meters with 1-minute time resolution. This allows grid conditions to be analyzed for their impact on al- gorithm performance and the level of assurance the algorithm would have when applied in a practical setting. Moreover, in their analysis, the authors implement the Spanish Low Voltage Distribution Network (LVDN) dataset known as POLA, which captures seven feeders’ informa- tion with more than 200,000 voltage readings at a time resolution of 15 minutes. This dataset is more realistic to the data received by utility companies’ smart meters. This work mainly aims the improvement of the reliability and applicability of phase identifi- cation methodologies. Further, the research appreciates the limitations of the envisaged MLV approach and endeavors into the possible social and ecological implications of the study, as well as broader reflections on machine learning within electrical engineering.