Buildings are one of the major sources of global energy consumption. In 2021, residential and commercial buildings were responsible for around 39% of total U.S. energy consumption and 74% of total U.S. electricity consumption. Consequently, research on the operation of building Heating Ventilation, and Air Conditioning (HVAC) systems can lead to significant energy savings and carbon emission reduction. Many control methods have been developed to provide solutions for building HVAC control problems, including model predictive control, nonlinear adaptive control, and decentralized control. However, most such approaches require detailed and exact building dynamics models, and an increase in the complexity of building dynamics would lead to significantly higher computational costs. 

One challenge of building research is the lack of a benchmark simulation environment for developing and evaluating different RL algorithms with realistic building models. In a recent work, we propose “BEAR” [1], a physics-principled Building Environment for Control and Reinforcement Learning. The platform allows researchers to benchmark both model-based and model-free controllers using a broad collection of standard building models in Python without co-simulation using external building simulators. We demonstrate the compatibility and performance of BEAR with different controllers, including both model predictive control (MPC) and several state-of-the-art RL methods with two case studies. BEAR is available at https://github.com/chz056/BEAR

In this project, we demonstrate how to bridge the gap between model accuracy and control tractability faced by neural networks, by explicitly constructing networks that are convex with respect to their inputs [2]. We show that these input convex neural networks (ICNNs) can be trained to obtain accurate models of complex physical systems. 

In particular, we design input convex recurrent neural networks to capture the temporal behavior of dynamical systems. The optimal controllers can be achieved by solving a convex model predictive control problem. Experiment results demonstrate the good potential of the proposed ICNN-based approach in a variety of control applications. In particular, we show that in the MuJoCo locomotion tasks, we could achieve over 10% higher performance using 5× less time compared with the state-of-the-art model-based reinforcement learning method; and in the building HVAC control example, our method achieved up to 20% energy reduction compared with classic linear models

Modern buildings encompass complex dynamics of multiple electrical, mechanical, and control systems. One of the biggest hurdles in applying conventional model-based optimization and control methods to building energy management is the huge cost and effort of capturing diverse and temporally correlated dynamics. Here we propose an alternative approach that is model-free and data-driven. By utilizing a high volume of data coming from advanced sensors, we train deep Recurrent Neural Networks (RNN) which could accurately represent the operation’s temporal dynamics of building complexes [3]. The trained network is then directly fitted into a constrained optimization problem with finite horizons. By reformulating the constrained optimization as an unconstrained optimization problem, we use an iterative gradient descent method with momentum to find optimal control inputs. Simulation results demonstrate the proposed method’s improved performances over model-based approach on both building system modeling and control.

References:

[1] Chi Zhang, Yuanyuan Shi, Yize Chen, "BEAR: Physics-Principled Building Environment for Control and Reinforcement Learning,", accepted to 14th ACM International Conference on Future Energy Systems (ACM e-Energy 2023).

[2] Y. Chen, Y. Shi, and B. Zhang, “Optimal Control Via Neural Networks: A Convex Approach,” In International Conference on Learning Representations (ICLR), 2019.

[3] Y. Chen, Y. Shi, and B. Zhang, “Modeling and optimization of complex building energy systems with deep neural networks”, Asilomar Conference on Signals, Systems, and Computers, 2017.