Technical Summary: BESTOpt – A Modular, Physics-Informed Machine Learning Framework for Building Modeling, Control, and Optimization

1. Problem Statement

The building sector faces an urgent imperative to decarbonize, requiring the decarbonization of approximately 10,000 buildings daily in the U.S. over the next 25 years. Achieving this requires a holistic framework capable of modeling, controlling, and optimizing highly interconnected ecosystems involving occupants, buildings, HVAC systems, distributed energy resources (DERs), and power grids.

Current tools suffer from three critical research gaps that prevent the realization of such a holistic framework:

• Scalability vs. Physical Consistency: Physics-based models (e.g., EnergyPlus, Modelica) offer high fidelity but are often too complex, requiring detailed metadata and manual calibration, making them impractical for large-scale deployment. Conversely, data-driven methods (e.g., standard machine learning) are scalable but often lack physical consistency and generalization capabilities under unseen conditions or limited/noisy data.
• Lack of Coordination and Interaction Modeling: Existing platforms (e.g., CityLearn, REopt) often treat buildings as static nodes with pre-defined load profiles, neglecting real-time dynamic interactions between indoor conditions, control systems, and DERs. They typically fail to model cross-domain interactions or centralized coordination across building clusters.
• Limited Transparency and Non-Differentiable Architectures: Most platforms are opaque, restricting user access to internal components (e.g., specific HVAC dynamics). This lack of modularity and differentiability hinders the integration of advanced gradient-based optimization techniques (like Model Predictive Control) and makes it difficult to interpret learned policies or ensure safe operation in model-free approaches like Reinforcement Learning.

2. Methodology

The authors propose BESTOpt, a modular, Physics-Informed Machine Learning (PIML) runtime environment designed to unify building applications including benchmarking, diagnostics, control, and optimization.

2.1 Hierarchical Architecture

BESTOpt adopts a four-level hierarchical structure to manage coordination across scales and domains:

1. Cluster Level: The topmost level representing the simulation scope (e.g., a campus or city), coordinating interactions among multiple domains and managing shared resources.
2. Domain Level: Defines energy flow types (currently thermal, electrical, and water).
3. System/Building Level: Represents packaged entities such as energy suppliers (systems) and demanders (buildings). A single system can influence multiple domains (e.g., an HVAC system consuming electricity while producing thermal output).
4. Component Level: The fundamental units (e.g., fans, coils, pumps, batteries, PV panels) used to construct systems.

2.2 Standardized Data Typology

To ensure interoperability, the framework standardizes data flow into four categories across all hierarchy levels:

• State: Internal, time-varying conditions (e.g., zone temperature, battery SOC).
• Action: Control variables sent by controllers (e.g., setpoints).
• Disturbance: Exogenous inputs not directly controlled (e.g., weather, occupancy, pricing).
• Observation: Data collected via sensors, inferred from models, or aggregated for monitoring.

Data flows directionally: higher-level actions are decomposed and dispatched downward, while lower-level states and observations are aggregated upward.

2.3 Module Design

The framework consists of three primary module types, all inheriting from a shared base class with Initialize(), Reset(), and Step() functions:

• Disturbance Modules: Generate time-varying inputs (weather, occupancy, prices). These can be static profiles, online API data, or dynamically forecasted using a CNN-LSTM-Bayesian neural network.
• Dynamic Modules: Represent physical systems.
  • Building Thermal Dynamics: Modeled using PI-ModNN (Physics-Informed Modularized Neural Network). This approach decomposes heat transfer into separate neural network modules with physics-informed constraints, ensuring physical consistency and generalization.
  • HVAC Systems: Aggregations of sub-components (fans, pumps, coils, chillers, etc.) ranging from empirical formulations to learning-based representations.
  • DER Systems: Includes PV panels, batteries, and EVs, implemented with equation-based formulations for interpretability and efficiency.
• Controller Modules: Generate control actions at various levels (Cluster, Domain, System, Component). The architecture supports rule-based logic, Model Predictive Control (MPC), and Reinforcement Learning (RL).

2.4 Differentiability and Modularity

A core innovation is the implementation of modules in a differentiable form. This allows seamless integration with gradient-based optimization techniques. The modular design allows independent updates, sharing, and replacement of components, ensuring transparency and flexibility.

3. Key Contributions

• PIML-Based Module Development: The integration of fundamental physical principles (conservation laws, symmetries) with machine learning to create modules that generalize well to unseen scenarios while maintaining physical consistency.
• Hierarchical Cross-Domain Architecture: A structured framework enabling multi-scale (component to cluster) and multi-domain (thermal, electrical, water) coordination, facilitating the modeling of complex interactions like HVAC-DER-grid co-optimization.
• Modular, Differentiable Design: A transparent architecture that supports plug-and-play development and enables advanced control strategies (MPC, RL) by providing explicit, differentiable system representations.

4. Results and Case Studies

The paper presents several case studies demonstrating BESTOpt's capabilities:

• Component-Level Control: A simulation of a single-zone building with a Fan Coil Unit (FCU) showed that BESTOpt accurately captures the mismatch between control commands and real system feedback (e.g., VFD vs. constant-speed fans), unlike platforms relying on predefined curves.
• Thermal Dynamic Performance: In a scenario where the HVAC system was turned off (an unseen condition), a standard LSTM baseline failed to predict the subsequent thermal drift, significantly underestimating zone temperatures. In contrast, the PI-ModNN model maintained physical consistency and accurately predicted the temperature rise, demonstrating superior generalization.
• Single-Family House with DERs: A simulation of a house with an FCU, PV, battery, and two EVs demonstrated the framework's ability to simulate fine-grained interactions between building dynamics, HVAC control, and DER coordination under rule-based strategies (e.g., Time-of-Use pricing).
• Cluster-Scale Simulation: A two-day simulation of a five-building residential cluster showed that PI-ModNN models trained on just three months of data (in under five minutes per building) could successfully maintain thermal comfort across diverse building envelopes. The framework successfully aggregated individual loads to show cluster-level consumption and diversity.

5. Significance and Claims

The authors position BESTOpt as a foundational platform to accelerate interdisciplinary research toward smart, resilient, and decarbonized building ecosystems. Its significance lies in:

• Bridging the Gap: It addresses the trade-off between the scalability of data-driven methods and the physical consistency of physics-based models.
• Enabling Advanced Control: By providing differentiable, transparent, and modular components, it unlocks the potential for advanced optimization techniques (like MPC and RL) that were previously constrained by opaque or non-differentiable platforms.
• Scalability: The hierarchical structure supports applications ranging from single-component control to centralized coordination of building clusters.

The paper modestly notes current limitations, such as increased training times for multi-zone buildings and the complexity of high-dimensional observation spaces for cluster-level controllers, identifying these as targets for future optimization. The authors envision BESTOpt evolving into an open, extensible platform for large-scale benchmarking and evaluation.