Stripping off the implementation complexity of physics-based model predictive control for buildings via deep learning

Abstract

Over the past decade, model predictive control (MPC) has been considered as the most promising solution for intelligent building operation. Despite extensive effort, transfer of this technology into practice is hampered by the need to obtain an accurate controller model with minimum effort, the need of expert knowledge to set it up, and the need of increased computational power and dedicated software to run it. A promising direction that tackles the last two problems was proposed by approximate explicit MPC where the optimal control policies are learned from MPC data via a suitable function approximator, e.g., a deep learning (DL) model. The main advantage of the proposed approach stems from simple evaluation at execution time leading to low computational footprints and easy deployment on embedded HW platforms. We present the energy savings potential of physics-based (also called 'white-box') MPC applied to an office building in Belgium. Moreover, we demonstrate how deep learning approximators can be used to cut the implementation and maintenance costs of MPC deployment without compromising performance. We also critically assess the presented approach by pointing out the major challenges and remaining open-research questions.

Full text

Stripping off the implementation complexity of
physics-based model predictive control for buildings
via deep learning
Ján Drgoˇ
na1,2,Lieve Helsen2,3, and Draguna L. Vrabie1
1Pacific Northwest National Laboratory, Richland, WA, USA
{jan.drgona, draguna.vrabie}@pnnl.gov
2Department of Mechanical Engineering, KU Leuven, Belgium
{jan.drgona, lieve.helsen}@kuleuven.be
3EnergyVille, Thor Park, Waterschei, Belgium
Abstract
Over the past decade, model predictive control (MPC) has been considered as
the most promising solution for intelligent building operation. Despite extensive
effort, transfer of this technology into practice is hampered by the need to obtain
an accurate controller model with minimum effort, the need of expert knowledge
to set it up, and the need of increased computational power and dedicated software
to run it. A promising direction that tackles the last two problems was proposed
by approximate explicit MPC where the optimal control policies are learned from
MPC data via a suitable function approximator, e.g., a deep learning (DL) model.
The main advantage of the proposed approach stems from simple evaluation at
execution time leading to low computational footprints and easy deployment on
embedded HW platforms. We present the energy savings potential of physics-
based (also called ’white-box’) MPC applied to an office building in Belgium.
Moreover, we demonstrate how deep learning approximators can be used to cut the
implementation and maintenance costs of MPC deployment without compromising
performance. We also critically assess the presented approach by pointing out the
major challenges and remaining open-research questions.
1 Introduction
Nowadays buildings use roughly
40 %
of the global energy (approx. 64 PWh), a large portion of
which is being used for heating, cooling, ventilation, and air-conditioning (HVAC) [
1
]. The energy
efficiency of buildings is thus one of the priorities to sustainably address the increased energy demands
and reduction of CO2emissions in the long term [2].
It has been shown that smart control strategies like model predictive control (MPC) can maximize
system-level efficiency for existing built environments, thus reducing the emissions of greenhouse
gases, and can improve the thermal comfort of the occupants, with reported energy use reductions of
15 % up to 50 % [3, 4, 5].
Despite this, the practical implementations of MPC are hampered by the challenge of obtaining an
accurate controller model with minimum effort, the need of expert knowledge to set it up, and the
need of increased computational power and dedicated software to run it [
6
]. Every building represent
a unique system which requires tailored modeling and control design.
33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.

The MPC in this work is based on detailed physical modeling of a real-life office building which
provides an accurate prediction of the building’s thermal behavior and high control performance.
On the other hand, the disadvantage of such high-fidelity MPC approach lies in its computational
demands and software dependencies. Here we are exploring the use of DL to learn the optimal
control policies from MPC data. The main advantage of the proposed method stems from its low
computational footprint, minimal software dependencies and easy deployment even on low-level
hardware without compromising control performance. The advantage compared to reinforcement
learning is its sample efficiency, because policies are learned with supervision from pre-computed
optimal control trajectories in realistic operational scenarios.
2 Optimization-based model predictive control using physical models
The office building considered in the experimental and simulation study called Hollandsch Huys is
located in Hasselt, Belgium. Hollandsch Huys represents a so-called GEOTABS building with slow
dynamics and complex heating ventilation and air conditioning (HVAC) system [
7
]. The building’s
layout consists of five floors divided into
12
thermal zones. For detailed description of the building
and physics-based modeling in Modelica language we refer to [
8
]. The main advantage of such a
high fidelity physics-based "digital twin" model stems from its potentially high prediction accuracy,
interpretability, and reliability. Based on the approach described in [
9
] the physics-based model can
be transformed to state-space representation with
700
states
x
,
301
disturbance signals
d
,
12
thermal
zones yand 20 control inputs u.
Fig. 1 shows the corresponding control configuration. The optimization-based MPC (OB-MPC)
computes the optimal control actions
u
, based on estimated states
x
via Kalman Filter (KF), for details
see [
10
,
11
]. The MPC problem is solved using a state-of the art optimization solver Gurobi [
12
]
running in the MATLAB environment. The non-linear weather forecaster model is running in
the Dymola environment and computes the forecasts of disturbances
d
(weather, occupancy), and
reference
r
trajectories based on actual weather data
w
obtained from the Dark Sky API [
13
]. Optimal
control actions at the current time step
u0
represent the heat flows to be delivered to the building and
are re-computed once per sampling time in so-called receding horizon control (RHC) fashion.
Hollandsch Huys
office building
u
y
d
r
w
Dark Sky
weather
y
ru
w
train replace
MPC & KF
OB-MPC
DL-MPC
Disturbances
and comfort
forecaster
Figure 1: Optimization-based MPC methodology with deep learning-based policy approximator.
3 Deep learning-based approximation of MPC policies
The central idea here is based on learning the optimal control policies from optimal trajectories
generated by OB-MPC via deep learning model in an imitation learning fashion, as shown in Fig. 1.
A detailed description of the applied methodology can be found in [
10
]. After training, the DL-MPC
policy replaces the computationally heavy and costly OB-MPC implementation. We use MATLAB’s
neural network toolbox for the design and training of the three-layer time-delayed neural network on
330 days of simulated operation of the original OB-MPC.

4 Experimental and simulation results
The real operational performance of the physics-based OB-MPC is compared to the conventional
rule-based controller (RBC) on a dataset of
72
days (
31
for MPC,
41
for RBC) during the transient
season (intermediate between spring and summer). The mean ambient temperature for the MPC
dataset is
17.3◦C
, and for RBC it is
18.8◦C
. The corresponding HP energy savings of OB-MPC are
equal to
50.4 %
, with a thermal comfort improvement of
50.5 %
. However, it is essential to mention
that that these are preliminary results for the transient season, that can not be generalized over all
seasons. Nevertheless, these results are encouraging and provide a glimpse of the energy-saving
potential of the proposed physics-based predictive control strategy in a real setting.
Subsequently, we evaluate the control performance on a simulated
30
-days test set, together with
the deployment cost reduction of the proposed DL-MPC with respect to OB-MPC. The simulation
setup is idealized, as no uncertainty in the feature space of both OB-MPC and DL-MPC is considered.
As a result, DL-MPC kept very high comfort satisfaction close to
100 %
, but it slightly increased
the energy use roughly by
3 %
w.r.t. high-fidelity OB-MPC. Yet, DL-MPC kept high energy saving
potential compared to the classical RBC. However, in contrast to the runtime and deployment cost
of OB-MPC, the presented neural policies require only a fraction of computational and memory
resources without the need for expensive software dependencies. In this case, we observed that
DL-MPC is roughly
50 000
-times faster and consumes
638
-times less memory. The overall control
performance, average CPU evaluation time per sample
1
, memory footprint
2
, together with the cost
associated with commercial software licenses 3are summarized in Tab. 1.
Table 1: Comparison of OB-MPC and DL-MPC. Performance indicators: simulation performance
on 30-days test set, computational and memory footprint, and software deployment cost.
Method Discomfort Energy use CPU time Memory SW Deployment
[K h] [kW h] [1×10−3s] [MB] Cost [$]
OB-MPC 0.0 801.2 26 843 415 18,050
DL-MPC 0.15 824.5 0.528 0.65 0
5 Conclusions, challenges and future work
In this work, we demonstrated the preliminary results of the energy-saving potential of the
optimization-based model predictive control (OB-MPC) based on a physical model in the oper-
ation of the real office building in Belgium. Additionally, we showed on simulation results, how deep
learning technology could be used to reduce the deployment cost of such advanced control strategies,
maintaining high control performance, while using only a fraction of computational resources.
However, several open-research problems remain unanswered. For example, what is the optimal
topology and hyperparameter setup for efficient representation of such problems? How to guarantee
satisfactory control performance far from the optimal trajectory? How sensitive is the policy to
uncertainty in weather forecast? Does the policy stabilize the closed-loop system? How to explicitly
include constraint handling properties of OB-MPC into DL-MPC policies? How can we use predictive
models and state estimation algorithms to further improve policy performance based on feedback?
How can we verify the policies using physics-based models? Can we parametrize the policies based
on physical parameters of the buildings to be used in a transfer learning fashion? Can we create
synthetic training datasets using generative models with the aid of physics-based modeling? Can we
use generative models to synthesize the policies directly from the building parameters?
Future work of the authors, includes deployment of trained DL-MPC policies in a real office building.
As the step towards computationally efficient and interpretable neural network policies for real-world
systems, the authors are focusing on the development of novel deep neural topologies inspired by the
sparse structure of the physics-based models and optimal control problems.
1
In case of OB-MPC the average runtime is the sum of
24.534 s
for the non-linear weather forecaster model
running in Dymola and 2.309 s for the MPC solution via Gurobi.
2
In case of OB-MPC, only the implementation code and actively used libraries are evaluated. We are omitting
the memory requirements of the MATLAB and the Dymola environments themselves.
3
Overall costs are computed as aggregate cost of MATLAB perpetual license (
2,150
$), Gurobi single user
license (10,000 $), and Dymola standard license (5,900 $).

References
[1]
IEA International Energy Agency and International Partnership for Energy Efficiency Cooperation. Build-
ing energy performance metrics - supporting energy efficiency progress in major economies. Technical
report, IEA Publications, 2015.
[2]
David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran,
Andrew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra
Luccioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Körding, Carla Gomes,
Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio.
Tackling climate change with machine learning. CoRR, abs/1906.05433, 2019.
[3]
Jan Širok
`
y, Frauke Oldewurtel, Jiˇ
rí Cigler, and Samuel Prívara. Experimental analysis of model predictive
control for an energy efficient building heating system. Applied Energy, 88(9):3079–3087, 2011.
[4]
D. Sturzenegger, D. Gyalistras, M. Morari, and R. S. Smith. Model predictive climate control of a swiss
office building: Implementation, results, and cost–benefit analysis. IEEE Transactions on Control Systems
Technology, 24(1):1–12, Jan 2016.
[5]
Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, and P. Haves. Model predictive control for the
operation of building cooling systems. IEEE Transactions on Control Systems Technology, 20(3):796–803,
2012.
[6]
J. Cigler, D. Gyalistras, J. Široký, V. Tiet, and L. Ferkl. Beyond Theory: the Challenge of Implementing
Model Predictive Control in Buildings. In Proceedings of 11th Rehva World Congress, Clima, 2013.
[7]
Elisa Van Kenhove, Jelle Laverge, Wim Boydens, and Arnold Janssens. Design Optimization of a
GEOTABS Office Building. Energy Procedia, 78:2989 – 2994, 2015. 6th International Building Physics
Conference, IBPC 2015.
[8]
D. Picard. Modeling, Optimal Control and HVAC Design of Large Buildings using Ground Source Heat
Pump Systems, PhD Thesis, KU Leuven, Belgium. 2017.
[9]
Damien Picard, Filip Jorissen, and Lieve Helsen. Methodology for obtaining linear state space building
energy simulation models. In Proceedings of the 11th International Modelica Conference, pages 51–58,
Paris, France, 2015.
[10]
J. Drgoˇ
na, D. Picard, M. Kvasnica, and L. Helsen. Approximate model predictive building control via
machine learning. Applied Energy, 218:199 – 216, 2018.
[11]
D. Picard, J. Drgoˇ
na, M. Kvasnica, and L. Helsen. Impact of the controller model complexity on model
predictive control performance for buildings. Energy and Buildings, 152:739 – 751, 2017.
[12] Inc. Gurobi Optimization. Gurobi optimizer reference manual, 2012.
[13] LLC. The Dark Sky Company. Dark Sky API.