Abstract: A predictive modeling approach to managing cellular network infrastructure is disclosed. In an embodiment, a method can include receiving raw data from a plurality of data sources populated while operating a cellular network. The method can then generate per-logical cell site data by normalizing the raw data based on a set of LCSs in the cellular network to generate per-LCS data. The method can then generate an example from the per-LCS data and generate a predicted energy consumption value for the given LCS by inputting the example into a predictive model (e.g., a decision tree-based model, such as an XGBoost model). From this output, the method can determine if the predicted energy consumption value is higher than an expected energy consumption value (e.g., a historical range of consumption). If so, the method can then label the given LCS as an outlier.