🌀 nonlinear sciences

Scalable Physics-Informed Neural Differential Equations and Data-Driven Algorithms for HVAC Systems

This paper presents a scalable, data-driven simulation framework for large-scale HVAC systems that combines physics-informed neural ordinary differential equations (PINODEs) with differential-algebraic equation (DAE) solvers and Bayesian optimization to achieve multi-fold speedups over high-fidelity models while maintaining low prediction errors and enforcing physical constraints.

Original authors: Hanfeng Zhai, Hongtao Qiao, Hassan Mansour, Christopher Laughman

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanfeng Zhai, Hongtao Qiao, Hassan Mansour, Christopher Laughman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your home's heating and air conditioning system (HVAC) as a massive, complex orchestra. It has many instruments (compressors, pipes, valves, heat exchangers) all playing together. To make the music sound perfect, the conductor needs to know exactly how every instrument will react to a change in the score, instantly and without a single wrong note.

For a long time, engineers had two choices for conducting this orchestra:

The "Super-Computer" Method: They used incredibly detailed physics models that knew every tiny vibration of every instrument. This was accurate, but it took so long to calculate that by the time the computer finished the math, the room was already too hot or too cold. It was like trying to conduct a symphony by calculating the physics of every single air molecule.
The "Guesstimate" Method: They used simple, fast shortcuts. These were quick, but they often got the physics wrong, leading to the system running inefficiently or breaking down over time.

This paper introduces a new, hybrid conductor that gets the best of both worlds: it's fast like a shortcut but accurate like a super-computer. Here is how it works, broken down into simple parts.

1. The "Smart Student" (PINODEs)

First, the researchers taught a computer program (a Neural Network) to act like a "smart student" for each individual part of the HVAC system, like a heat exchanger (the part that cools or heats the air).

The Old Way: Usually, these students were just memorizing answers. If you asked them a slightly different question, they might guess wrong.
The New Way (PINODE): The researchers taught the student the laws of physics (like "energy cannot be created or destroyed"). Instead of just memorizing data, the student learned the rules of the game.
- Analogy: Imagine teaching a child to drive. Instead of just memorizing the route to the store, you teach them the rules of the road (stop at red lights, yield to pedestrians). Now, no matter what route they take, they know how to drive safely.
- The Magic Trick: The model predicts how much "stuff" (refrigerant mass and energy) is inside the pipes. Because it knows the rules, it can predict how that stuff changes over time without getting confused, even over long periods.

2. The "Team Captain" (The Corrector Network)

Even smart students make small mistakes when they work together in a big group. If you have 32 of these "smart students" trying to play an orchestra, their tiny individual errors can add up to a huge mess by the end of the song.

The Solution: The researchers added a "Team Captain" (a tiny, lightweight neural network).
How it works: The Captain watches the first few minutes of the orchestra playing. It notices, "Hey, the second violin is always slightly flat," or "The bass drum is hitting a beat too early."
The Fix: The Captain then gently nudges the whole group to stay in tune for the rest of the song. It doesn't need to relearn everything; it just learns to fix the specific mistakes the group makes when they interact.

3. The "Traffic Cop" (DAE Solvers)

In a real HVAC system, everything is connected. If one pipe changes pressure, it affects the pipes connected to it immediately. It's like a traffic intersection where cars from four different directions have to merge without crashing.

The Problem: Standard computer math struggles with these "instant connections." It tries to solve them one by one, which causes traffic jams (math errors) or crashes.
The Solution: The researchers used a special type of math solver (called a DAE solver) that acts like a super-smart Traffic Cop.
- Instead of letting cars go one by one, the Traffic Cop looks at the whole intersection at once. It ensures that the pressure is equal everywhere and that the mass of air flowing in equals the mass flowing out.
- They also used a "Bayesian Optimization" tool, which is like a GPS for tuning. Instead of guessing the best settings for the Traffic Cop, the GPS automatically tries thousands of combinations to find the perfect balance between "driving fast" (speed) and "driving safely" (accuracy).

4. The Results: Fast, Accurate, and Scalable

The team tested this new system on everything from a simple two-compressor setup to a massive system with 32 compressors (like a huge office building or a data center).

Speed: The new system is 4 to 9 times faster than the old, high-fidelity physics simulators. It can predict what will happen in the next hour in less than a minute.
Accuracy: Despite being so fast, it is incredibly accurate. The error rate is less than 2.5%. That's like predicting the weather for next week and being right about the temperature within a degree or two.
Scalability: It works just as well for a small house as it does for a massive industrial complex.

Why Does This Matter?

Think of this technology as giving building managers a crystal ball.

Before: They had to guess how to run their AC to save money, or they had to wait hours for a computer to tell them the best setting.
Now: They can run thousands of "what-if" scenarios in seconds. They can say, "What if the temperature outside spikes by 10 degrees? What if we turn off the AC in the east wing?" The system tells them the answer instantly and accurately.

This means buildings can use less energy, save money, and be more comfortable, all because we finally taught computers to understand the physics of air conditioning without getting bogged down in the math.

1. Problem Statement

Heating, Ventilation, and Air Conditioning (HVAC) systems are major energy consumers, yet their accurate modeling and real-time control are hindered by a fundamental trade-off:

High-Fidelity Physics-Based Models: Tools like Modelica or EnergyPlus offer high accuracy by solving complex, stiff, nonlinear Differential-Algebraic Equations (DAEs) that couple mass, energy, and momentum conservation. However, they are computationally expensive, making them unsuitable for real-time control, optimization, or large-scale co-simulation.
Data-Driven Surrogates: Standard machine learning models (e.g., LSTMs, standard Neural ODEs) are computationally efficient but often violate physical conservation laws (mass/energy), leading to unphysical predictions and error accumulation over long time horizons. They also struggle with the numerical stability required for stiff thermo-fluid dynamics.

The core challenge is to develop a surrogate model that matches the accuracy of high-fidelity physics simulators while achieving the computational speed of data-driven methods, all while strictly adhering to thermodynamic principles.

2. Methodology

The authors propose a hybrid framework that integrates Physics-Informed Neural Ordinary Differential Equations (PINODEs) at the component level with DAE solvers at the system level, augmented by a Corrector Network.

A. Component Level: Implicit PINODE Formulation

Instead of treating conserved quantities as inputs, the model treats them as outputs.

Architecture: An Encoder-Neural ODE-Decoder structure.
- Encoder: A Gated Recurrent Unit (GRU) processes historical boundary conditions (air/refrigerant properties) to extract a latent representation.
- Neural ODE: A continuous-time latent dynamics model (using a gated architecture similar to GRUs and Layer Normalization) evolves the state. This ensures stable gradient flow and prevents vanishing/exploding gradients during long-horizon integration.
- Decoder: A GRU maps the latent state back to observable outputs.
Implicit Output: The model predicts the refrigerant mass ( $M_r$ ) and internal energy ( $E_{hx}$ ) directly.
Physics-Informed Training:
- The model computes time derivatives ( $\dot{M}_r, \dot{E}_{hx}$ ) via automatic differentiation.
- A physics loss term ( $L_{phys}$ ) enforces consistency between these predicted rates and the true rates derived from mass and energy balance equations.
- A conservation loss ( $L_{cons}$ ) ensures the predicted state variables match ground truth.

B. System Level: DAE Solver Integration

Large-scale HVAC networks involve algebraic constraints (pressure equilibrium at junctions, mass flow consistency) that standard ODE solvers cannot handle.

DAE Formulation: The system is formulated as an index-1 DAE where differential variables are $M_r$ and $E_{hx}$ , and algebraic variables are junction pressures.
Solvers: The framework integrates two robust DAE solvers:
- IDA: Uses modified Newton iteration without explicit Jacobian formation.
- DASSL: Explicitly forms a step-dependent Jacobian, offering better convergence for stiff systems.
Constraint Enforcement: At each time step, the solver explicitly solves for junction pressures to satisfy mass flow balance before advancing the differential states.

C. System-Level Correction

To address residual errors caused by algebraic coupling and numerical integration:

Corrector Network: A lightweight neural network is trained on short trajectory segments (initial 850 steps) to learn systematic biases between the PINODE-DAE predictions and high-fidelity benchmarks (Dymola).
Deployment: The network outputs correction terms for mass and energy, which are smoothed using Gaussian Processes (GP) and Exponential Moving Averages (EMA) to ensure temporal stability before being applied to the full simulation.

D. Automated Parameter Tuning

Bayesian Optimization: Used to automatically tune solver parameters (tolerances, step sizes, output intervals) for Algebraic, IDA, and DASSL solvers. This balances the trade-off between accuracy (MAPE) and computational efficiency (simulation time).

3. Key Contributions

Implicit PINODE Formulation: A novel approach where conserved quantities ( $M_r, E_{hx}$ ) are model outputs, enabling direct enforcement of conservation laws via automatic differentiation during training.
Stable Long-Horizon Dynamics: The use of gated architectures and layer normalization within the Neural ODE ensures numerical stability for stiff thermo-fluid systems over long prediction horizons.
Explicit DAE System Integration: A unified framework that couples learned component dynamics with DAE solvers (IDA/DASSL) to explicitly enforce algebraic constraints (pressure/mass flow), ensuring thermodynamic validity.
Scalability: The framework successfully scales to HVAC systems with up to 32 compressor–condenser pairs, a significant leap from previous data-driven attempts.
Automated Optimization: The use of Bayesian optimization to systematically tune solver parameters across different solver types, identifying Pareto-optimal configurations for accuracy vs. speed.

4. Results

Accuracy: The proposed framework achieves a Mean Absolute Percentage Error (MAPE) below 2.5% (best case: 2.04%) compared to high-fidelity Dymola simulations.
Speed: The method achieves a 4–9× computational speedup over high-fidelity simulators.
- DAE-IDA Solver: 58.91s simulation time (2.04% MAPE) vs. ~146.8s for the baseline.
- Fastest Configuration: 15.43s simulation time (3.56% MAPE), representing a ~9.5× speedup.
Scalability: The system successfully simulated configurations up to $N_c = 32$ (32 compressors/condensers). The DAE-IDA solver demonstrated the most favorable scaling behavior, maintaining moderate runtimes even as system size increased, whereas the algebraic solver showed steeper computational growth.
Corrector Effectiveness: The corrector network significantly reduced systematic biases in mass and energy predictions, tightening the parity plot clustering around the diagonal and improving transient tracking.

5. Significance

This work bridges the critical gap between physical accuracy and computational efficiency in HVAC modeling. By embedding conservation laws directly into the neural network's latent dynamics and using robust DAE solvers for system constraints, the framework overcomes the limitations of purely data-driven models (unphysical results) and purely physics-based models (computational intractability).

The ability to simulate large-scale, multi-compressor HVAC systems in near real-time with high fidelity opens new avenues for:

Real-time Control: Enabling Model Predictive Control (MPC) for complex building energy systems.
System Optimization: Rapidly exploring design spaces for large commercial HVAC installations.
Fault Detection: Providing accurate baselines for anomaly detection in real-world operations.

The paper establishes a new standard for "thermodynamically consistent" data-driven surrogates, proving that neural networks can be rigorously constrained by physics to solve industrial-scale engineering problems.