Electricity demand in Yucatán is rising, making energy efficiency in public buildings a policy priority. We propose a hybrid, unsupervised pipeline that combines K-means (global typologies) with DBSCAN (density-based anomaly discovery) to segment ~350 public buildings by energy behavior. After systematic preprocessing (outlier control via IQR, multivariate imputation, Min–Max scaling), K-means revealed four consistent energy typologies. DBSCAN, tuned with a k-dist heuristic, flagged >60% of records as atypical patterns that merit targeted audits. Using an energy-use-intensity benchmark of 120 kWh m-2 yr.-¹, the high-intensity cluster (n ≈ 65; EUI ≈ 175 kWh m-2 yr-¹) shows an ≈20–25% reduction potential relative to current levels, supporting the prioritization of retrofits and operational measures. PCA projections and georeferenced maps aid interpretation and policy targeting. The results indicate that the hybrid approach captures both dominant patterns and irregular behaviors, translating raw meter data into actionable, scalable guidance for evidence-based public programs in building efficiency.