Physics-informed neural ordinary differential equations for multi-zone residential thermal modeling

• Proposed NODEs match predictive accuracy of traditional NODEs and LSTMs. • IG attributions for proposed NODEs align with localized heating and solar gains. • A masked architecture improves interpretability by incorporating physical priors. • Proposed NODEs are robust to predicting unevenly spaced data from occupied residences. Heating, Ventilation and Air Conditioning (HVAC) systems account for a large share of residential energy use and drive peak electricity demand. Yet, deploying smart, adaptive temperature controllers is hindered by the cost and complexity of obtaining accurate building models. In this work, a physics‑informed neural differential equation (NODE) approach is introduced to learn multi‑zone thermal dynamics of electrical baseboard heated homes from thermostat data. First, a fully connected data‑driven NODE captures temperature evolution and heating input effects without any prior building specifics. Second, insights of the thermal processes are embedded into the network via a masked architecture, improving interpretability and reducing non-physical cross‑zone couplings. Both NODE variants are benchmarked against a discrete‑time LSTM on hourly experimental data from a controlled test house and then the masked NODE architecture is used on irregularly sampled measurements from 16 occupied Québec residences. Experiments show that (a) masked NODEs match predictive accuracy of fully connected NODEs and LSTMs with respective average RMSEs of 0.49 °C, 0.40 °C and 0.55 °C on the data from the controlled test house, (b) integrated‑gradient attributions align with physical heat transfer laws in scenarios involving localized heating and solar gains, and (c) NODE models remain robust to unevenly spaced data with an average weighted RMSE of 0.48 °C over the 16 occupied residences. These results demonstrate that NODEs offer a broadly adaptable and interpretable framework for residential HVAC modeling.