A hybrid modeling approach to predicting HVAC demand in Japanese houses using physics-based simulation and Artificial Neural Networks

Abstract

Understanding the relationship between energy usage patterns and consumption is critical for improving building energy efficiency through accurate forecasting. While numerous data-driven models have been trained on historical energy data to enhance prediction, few have incorporated occupant-related parameters. This study proposes an alternative approach for estimating energy use by integrating detailed household data, including occupancy, HVAC setpoint, building characteristics, and weather conditions. Four predictive models were developed: (1) a physics-based model via EnergyPlus, (2) a standalone data-driven machine learning (ML) model, (3) an ML model excluding setpoint data, and (4) a hybrid model integrating EnergyPlus with ML Artificial Neural Networks modeling. Simulation results demonstrate the superiority of the hybrid approach, emphasizing the vital role of air conditioning setpoint data in improving hourly air conditioning load prediction accuracy for individual residential units. By integrating physics-based and data-driven methods, this framework captures specific energy-use patterns in small-scale housing and provides actionable energy benchmarking and efficiency recommendations for residential communities.