House Thermal Model Estimation: Robustness Across Seasons and Setpoints

TL;DR

This work systematically compares three parameter-estimation methods—Nonlinear Least Squares, Batch Estimation, and Maximum Likelihood Estimation—for RC-network based house thermal models, evaluating both a four-state physics-grounded model (SM4) and two reduced-order models (SM2 and SM1). Through forward simulation across four seasons and multiple HVAC setpoints, it demonstrates that BE with SM-1 offers the best balance of accuracy and robustness for predicting indoor temperature $T_z$ and HVAC power $P_{HVAC}$, while $T_z$ remains robust to seasonal and setpoint changes and $P_{HVAC}$ is more sensitive to conditioning. The study provides practical guidance on selecting estimation techniques, model order, and training data for grid-edge demand response applications. Overall, the results support using reduced-order RC models with probabilistic estimation (BE/MLE) to enable scalable, robust house thermal modeling under varying operating conditions. The work lays groundwork for deploying RC-network models in real houses and large-scale grid-edge simulations.

Abstract

Achieving the flexibility from house heating, cooling, and ventilation systems (HVAC) has the potential to enable large-scale demand response by aggregating HVAC load adjustments across many homes. This demand response strategy helps distribution grid to flexibly ramp-up or ramp-down local load demand so that it can optimally match the bulk power system generation profile. However, achieving this capability requires house thermal models that are both computationally efficient and robust to operating conditions. In this work, parameters of the Resistance-Capacitance (RC) network thermal model for houses are estimated using three optimization algorithms: Nonlinear Least Squares (NLS), Batch Estimation (BE), and Maximum Likelihood Estimation (MLE). The resulting models are evaluated through a Forward-Simulation across four different seasons and three setpoints. The results illustrate a principled way of selecting reduced order models and estimation methods with respect to the robustness offered to seasonal and setpoint variations in training-testing datasets

Figures (2)

• Figure 1: Performance comparison of $T_{z}$ and $P_{HVAC}$ trajectories across different testing periods (spring, winter, fall) and setpoints (18°C, 26°C).
• Figure 2: Comparison of average accuracy for $T_{z}$ and $P_{HVAC}$ trajectories across training and testing periods.