INTEGRATED LOAD CLUSTERING AND MACHINE LEARNING FORECASTING FOR BUILDING ENERGY ANALYSIS

Abstract

Cities in hot-humid climates experience persistent cooling demand and rising electricity consumption, making accurate short-term load forecasting and systematic characterization of building energy use essential for operational energy management. This study presents a data-driven framework to derive representative load archetypes across building sectors and generate hourly electricity forecasts at high temporal resolution to support energy-related decision-making. One year of hourly electricity records from 24 buildings (8,760 observations per building) in Manama, Bahrain, is analyzed. Temporal and autoregressive features are derived from standardized timestamps, including Load(t-1), Load(t-2), Load(t-24), hour, day, and month. K-Means clustering is used to identify daily load archetypes, with the number of clusters selected using elbow and silhouette analyses. Three supervised machine-learning models, Random Forest (RF), Gradient Boosting (GB), and Artificial Neural Network (ANN), are trained to predict hourly electricity demand and evaluated using R² and RMSE. Three distinct consumption archetypes are identified: schedule-driven buildings, large commercial systems with sustained demand, and one atypical commercial educational facility. ANN achieved the highest hold-out accuracy (R² = 0.89, RMSE = 344 kW), whereas RF showed the strongest generalization under CV-5 (R² = 0.78, RMSE = 596 kW) and was therefore selected. The proposed clustering-forecasting framework demonstrates a scalable, model-agnostic approach for comparative energy analysis, demand-response prioritization, and operational energy optimization in electricity-intensive urban environments.