IoT Data-Driven Energy Prediction for Smart Buildings: A Comparative Analysis of Machine Learning Models

Smart buildings generate large volumes of high-frequency IoT data, which are often affected by noise and variability, making accurate energy forecasting challenging. This study investigates the impact of temporal data granularity on prediction performance by comparing three machine learning models: Linear Regression, Random Forest, and Gradient Boosting. A dual-scale evaluation is conducted using both hourly and daily aggregated datasets. The results reveal that higher data granularity does not necessarily lead to better predictive accuracy. Hourly models suffer from volatility and noise, while daily aggregation significantly improves stability and reduces prediction error. Specifically, Gradient Boosting achieves the best performance for hourly forecasting (R2=0.29), effectively capturing short-term fluctuations, whereas Random Forest performs better for daily prediction, achieving the lowest RMSE of 24.02 kWh. These findings demonstrate that temporal resolution plays a critical role in model performance and should be considered alongside algorithm selection. The study provides practical guidelines for selecting appropriate machine learning models based on the operational time horizon of smart building energy management systems.