November 20, 2026
Continuous-Time Models for HVAC Forecasting and Control
How continuous-time machine-learning models can make building energy management more scalable and effective. This article highlights how neural differential equations, especially Controlled Differential Equations, can forecast indoor temperature and HVAC power use from high-level metering data without requiring detailed building models or deep equipment access.
Summary from research presented in Neural differential equations for temperature control in buildings under demand response programs by Department of Mathematics and Industrial Engineering, Institut d ’intelligence artificielle de Quebec, Groupe d’etudes et de recherche en analyse des decisions and Universite de Montreal, in collaboration with Brainbox AI.
Key highlights
- Continuous-time neural differential equation models forecast indoor temperature and HVAC power consumption from metering data, reducing the need for detailed building-specific physics models
- A largely data-driven approach, enhanced with physics-informed structure, makes the models adaptable across buildings while improving interpretability
- Continuous-time models are more sample-efficient than discrete-time approaches and are more robust to missing or irregular observations
- Embedding the models in a model predictive control framework enables temperature setpoint optimization that lowers HVAC power during demand response events while balancing occupant comfort
- Results on simulated and real building data show that continuous-time latent-state models can support practical demand response control with fewer training samples than discrete-time counterparts
Machine learning for building energy management
Continuous-time machine learning models may improve HVAC forecasting and control in buildings by addressing a persistent challenge in building operations: HVAC systems are among the largest contributors to peak electricity demand, yet most control strategies require detailed building models, extensive engineering effort or deep access to equipment-level controls. That makes scalable energy management difficult across diverse building portfolios.
A more adaptable alternative is proposed using neural differential equations, especially Controlled Differential Equations (CDEs), to forecast indoor temperature and HVAC power consumption using only high-level metering data. Rather than relying on detailed physical descriptions of a building, the models are trained directly from observed data. This approach can generalize across diverse buildings while still capturing the continuous nature of thermal dynamics.
Continuous-time models are especially useful when building data is limited, where they generally outperform discrete-time baselines for HVAC power prediction while maintaining similar temperature forecast accuracy; performance is strongest when models operate in a latent state space rather than directly in observation space. Although embedding energy conservation principles into the model improves interpretability and preserves causal relationships between power and temperature, it does not consistently improve predictive accuracy, suggesting that well-designed continuous-time, data-driven models can perform strongly without heavily engineered physical priors.
Beyond forecasting, the models are shown to support model predictive control for demand response. The proposed control framework adjusts indoor temperature setpoints to keep HVAC power below a defined maximum during peak events while minimizing comfort impacts. This is especially relevant for grid-interactive buildings, where flexibility is becoming a more important part of energy demand management strategies.
Machine learning can move beyond forecasting alone and support actionable control strategies. The broader implication is that more flexible, data-driven control architectures may help buildings participate more effectively in demand response programs while reducing peak load and preserving occupant comfort.
Topic Tags
-
References:
Taboga, V., Gehring, C., Mathieu Le, C., Dagdougui, H., & Bacon, P.-L. (2024b). Neural differential equations for temperature control in buildings under demand response programs. Applied Energy, 368(0306–2619), 123433. https://doi.org/10.1016/j.apenergy.2024.123433
English