Core Concepts
A data-driven modeling and control framework is presented for physics-based building emulators, using differentiable surrogate models and optimization-based nonlinear control methods to enhance model evaluations, provide cost-effective gradients, and maintain good predictive accuracy for Model Predictive Control (MPC).
Abstract
The paper presents a modeling and control framework for building HVAC systems using differentiable surrogate models compatible with optimization-based nonlinear control methods. The approach consists of two key components:
Offline training of differentiable surrogate models (Linear, MLP, LSTM) that accelerate model evaluations, provide cost-effective gradients, and maintain good predictive accuracy for the receding horizon in MPC.
Formulating and solving nonlinear building HVAC MPC problems using gradient descent and sequential quadratic programming (SQP) methods, with the goal of minimizing energy consumption while maintaining occupant comfort constraints.
The framework is evaluated extensively using multiple surrogate models and optimization frameworks across various test cases available in the Building Optimization Testing Framework (BOPTEST). The results demonstrate the superiority of LSTM models in predictive accuracy for the single-zone case, while MLP models perform better in the multi-zone case. The control results show that the gradient-based approach with LSTM performs best in the single-zone case, while SQP with MLP achieves the lowest power consumption in the multi-zone case.
The modular and customizable nature of the framework allows it to be adapted to various control approaches and test cases, providing a path towards prototyping predictive controllers in large buildings with emphasis on scalability and robustness.
Stats
The total energy consumption and thermal discomfort are reported as key performance metrics.