Experimental Data-Driven Model Predictive Control of a Hospital HVAC System During Regular Use

Abstract

Herein we report a multi-zone, heating, ventilation and air-conditioning (HVAC) control case study of an industrial plant responsible for cooling a hospital surgery center. The adopted approach to guaranteeing thermal comfort and reducing electrical energy consumption is based on a statistical non-parametric, non-linear regression technique named Gaussian processes. Our study aimed at assessing the suitability of the aforementioned technique to learning the building dynamics and yielding models for our model predictive control (MPC) scheme. Experimental results gathered while the building was under regular use showcase the final controller performance while subject to a number of measured and unmeasured disturbances. Finally, we provide readers with practical details and recommendations on how to manage the computational complexity of the on-line optimization problem and obtain high-quality solutions from solvers.

Lessons Learned from Data-Driven Building Control Experiments: Contrasting Gaussian Process-based MPC, Bilevel DeePC, and Deep Reinforcement Learning

“…In order to go from learning 'static functions' f (x) to dynamical systems, one can employ GPs to build either autoregressive or state-space models (see e.g. [25] and [28]). We opted for the former and used the final expressions within our MPC formulation [25].…”

“…[25] and [28]). We opted for the former and used the final expressions within our MPC formulation [25]. Whereas the mean function µ(x) plays the role of the nominal dynamics, the variance information var(x) is used to tighten the optimization problem constraints, conferring on it a certain degree of robustness.…”

“…The building features an industrial forced-air cooling system to guarantee indoor thermal comfort. The study presented in [25] was concerned with its ophthalmology section, composed of three rooms: two operating theaters and one intermediate zone. 31 days worth of clean data were collected, sampled every 10 min, and three individual GP models were then learned based roughly on 275 points each.…”

“…In Fig. 2, the temperatures and actuation profiles throughout two days are shown, and the period highlighted in grey corresponds to a chiller fault discussed in [25]. The adopted GP-MPC solution was shown to yield good performance even when subject to harsh outdoor conditions and strong internal disturbances (see [25,Sec.…”

“…In this work, we focus our analysis on three real-world building control experiments using distinct combinations of models and controllers: GP-based MPC [25], bilevel DeePC [26] -relying on behavioral system theory -and DRL matched with NN dynamical models [27]. This selection is representative of the various possible choices as it covers linear, nonlinear, parametric, and non-parametric models, as well as both predictive and explicit control laws.…”