Residential and small commercial buildings remain central to urban electricity demand, yet appliance-level visibility is often limited by the cost and installation burden of intrusive sub-metering. This study presents a machine-learning framework for non-intrusive load monitoring (NILM) that recovers appliance-level consumption patterns from a single smart-meter stream using steady-state active power. The framework combines a Bagging regressor for load prediction with a modified multiclass K-Nearest Neighbour (KNN) classifier for appliance identification and is evaluated on the Dutch Residential Energy Dataset (DRED), a low-frequency 1 Hz dataset containing aggregate and appliance-level measurements from a monitored household in the Netherlands. Using a 70% training split and 30% testing split, zero-imputation for sparse missing values, and time-derived features capturing hourly and daily usage variation, the proposed pipeline achieves strong predictive performance while preserving computational efficiency suitable for smart-meter deployment. The Bagging regressor records the best regression performance among the evaluated models, with a mean squared error of 99.3152, root mean squared error of 9.9657, mean absolute error of 0.2813, and coefficient of determination of 0.9624. For appliance classification, the optimised multiclass KNN model achieves a precision of 0.7497, recall of 0.7825, F1-score of 0.7531, and overall accuracy of 78.25%. Descriptive analysis of the monitored period further shows consistent refrigerator demand, peak aggregate usage on 6 and 13 July, and the lowest household demand between midnight and 05:33. These results support the use of low-frequency NILM as a practical and scalable component of urban energy analytics, with direct relevance for smart-home feedback, demand-side management, infrastructure planning, and carbon-aware city operations.