Robust Energy-Constrained Frequency Reserves From Aggregations of Commercial Buildings

This article has been accepted for inclusion in a future issue of this journal.
Content is final as presented, with the exception of pagination.
IEEE TRANSACTIONS ON POWER SYSTEMS 1
Robust Energy-Constrained Frequency Reserves

From Aggregations of Commercial Buildings
Evangelos Vrettos, Student Member, IEEE, Frauke Oldewurtel, Member, IEEE, and Göran Andersson, Fellow, IEEE
Abstract—It has been shown that the heating, ventilation, The idea of using flexible loads to balance the power grid,
and air conditioning (HVAC) systems of commercial buildings which is known as demand response (DR), was already pro-
can offer ancillary services to power systems without loss of posed in the 1980's [2]. Nowadays, DR has become a very
comfort. In this paper, we propose a new control framework
for reliable scheduling and provision of frequency reserves by active field of research firstly, because the need for AS is
aggregations of commercial buildings. The framework incorpo- increasing due to larger RES shares [3] and secondly, due to
rates energy-constrained frequency signals, which are adopted by recent advancements in communication technologies. Loads
several transmission system operators for loads and storage de- with thermal storage are particularly suitable for DR because
vices. We use a hierarchical approach with three levels: 1) reserve their consumption can be shifted in time without impact on
capacities are allocated among buildings (e.g., on a daily basis)
using techniques from robust optimization; 2) a robust model consumer comfort.
predictive controller optimizes the HVAC system consumption One line of research has focused on DR with thermostatically
typically every 30 minutes; and 3) a feedback controller adjusts controlled loads (TCLs), i.e., small residential loads such as re-
the consumption to provide reserves in real time. We demonstrate frigerators, air conditioners, space and water heaters. The key
how the framework can be used to estimate the reserve capacities challenges here are modeling of a large aggregation of TCLs and
in simulations with typical Swiss office buildings and different
reserve product characteristics. Our results show that an aggre- control with limited communication between the controller and
gation of approximately 100 buildings suffices to meet the 5-MW the loads. Among others, [4]–[8] addressed these challenges,
minimum bid size of the Swiss reserve market. whereas [9] provided a state of the art in DR with TCLs up to
2013 including earlier works in this field.
Index Terms—Ancillary services, demand response, energy con-
straints, frequency control, robust model predictive control. A second line of research considered DR with the heating,
ventilation, and air conditioning (HVAC) systems of commer-
cial buildings, which is also the focus of this paper. HVAC sys-
I. INTRODUCTION tems are interesting for DR because they are large consumers
with high thermal inertia, and they usually include a building
automation and communication system that facilitates DR im-
I N many countries, large amounts of renewable energy
sources (RES) are integrated in the power system and they
increase the operational uncertainty due to their fluctuating
plementation, see for example [9] and references therein. The
role of information systems and automated model-based control
for energy efficiency and DR was discussed in [10]. Commer-
nature. In a power system, supply and demand of electric
cial HVAC systems are typically complex with many control
power must be balanced to keep frequency and voltage close
variables and cascaded loops, and thus low-order models are
to their nominal values. To maintain this balance, transmission
preferable for control purposes.
system operators (TSOs) procure ancillary services (AS), e.g.,
Many approaches have been proposed in the literature to
frequency and voltage control, traditionally from conventional
obtain simplified building models, which can be classified
generators. However, if properly aggregated and controlled,
into neural network models, linear parametric models, and
loads can provide AS without environmental impacts and,
lumped thermal resistance and capacitance (RC) models [11].
possibly, more efficiently and at a lower cost compared to
In contrast to neural network models, the parameters of linear
generators [1].
parametric and RC models can be associated with the building's
physical parameters. Simple two-state RC models, where one
Manuscript received January 31, 2015; revised June 14, 2015, October
state is the room air temperature and the other one is the temper-
05, 2015, and December 05, 2015; accepted December 16, 2015. The work
of E. Vrettos and G. Andersson was supported in part by Nano-Tera.ch for ature of a single lumped thermal mass, were already proposed
the HeatReserves project and the Swiss Federal Office of Energy and Swis- in the 1980's [12]. Subsequent work investigated the advantages
selectric Research for the “SmartGrid-Polysun: Design Tool for Local Load
of modeling a building using more than one temperature nodes
Management” project. The work of F. Oldewurtel was supported by the EU
FP7-PEOPLE-2011-IOF Programme under Grant 302255, Marie Curie project for each opaque building element [13], [14]. Different variants
“Stochastic Model Predictive Control, Energy Efficient Building Control, of linear parametric models, including Box-Jenkins models and
Smart Grid”. Paper no. TPWRS-00135-2015.
autoregressive moving average models with external inputs,
E. Vrettos and G. Andersson are with the EEH-Power Systems Laboratory,
Swiss Federal Institute of Technology (ETH) 8092 Zurich, Switzerland (e-mail: were compared in [15]. A very simple linear parametric model
[email protected]; [email protected]). for DR applications was proposed in [16] to model the change
F. Oldewurtel is with the Department of Electrical Engineering & Computer
in HVAC power consumption after a global thermostat reset.
Sciences, University of California, Berkeley, CA 94720-2284 USA (e-mail: old-
[email protected]). Most of the early work on commercial buildings investigated
Digital Object Identifier 10.1109/TPWRS.2015.2511541 the use of building thermal mass for load shifting and peak
0885-8950 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON POWER SYSTEMS
shedding (see [17] for an overview of relevant simulation and aggregation can determine the amount of reserves it can col-
experimental results up to 2003). For example, [18] investigated laboratively offer, and how the buildings should be controlled
in simulation how the zone temperature setpoints can be opti- to achieve reserve provision without compromising occupants'
mally varied to reduce peak demand and take advantage of low comfort. The main contributions of this paper are threefold.
nighttime electrical rates, whereas [19] performed a set of ex- The first contribution is the development of a new three-level
periments to demonstrate the potential for load shifting and load hierarchical control framework to address the reserve provi-
leveling with optimal control. Predictive optimal control was sioning problem. The goal of the first level is to estimate the
compared against conventional control strategies (chiller-pri- SFC reserve capacity that can be robustly offered by a building
ority and storage-priority control) in [20] considering the min- aggregation considering weather conditions, occupancy, elec-
imization of either energy charges or peak demand charge, and tricity and reserve prices, and the comfort zone. The second
in [21] considering real-time pricing. level is an MPC that minimizes energy consumption in each
Load shifting and peak load reduction have been also ad- building, while ensuring that the scheduled reserves can be pro-
dressed in more recent works. For example, [22] developed vided. In the third level, the incoming SFC signal is tracked
a model predictive controller (MPC) for a building cooling controlling the HVAC system's consumption using a feedback
system, which uses a simplified nonlinear model and relies on controller. The framework actively allocates reserves among the
a moving window blocking strategy to reduce computation buildings of the aggregation based on their individual character-
time. The authors of [23] designed an MPC that incorporates a istics, which is expected to maximize the reserve potential. This
time-varying tariff for end-consumers, which reflects the true is in contrast to [25], [26], [28] that estimated the capacity of a
marginal costs of electricity provision, and they investigated the group of buildings by simply scaling up the capacity of a single
potential for reduction of peak electricity demand in buildings. building.
Recently, some papers have investigated provision of “fast” The second contribution is modeling the SFC signal as uncer-
AS from commercial buildings, in particular secondary fre- tain and handling reserve biases. Bias is the SFC signal's inte-
quency control (SFC), which is also known as automatic gral, i.e., the amount of energy that is withdrawn from or stored
generation control (AGC), load frequency control (LFC), to the buildings over time. Ideally, the SFC signal would be ap-
automatic frequency restoration reserve (FRR), or regulation proximately zero-energy and any biases would be absorbed by
service. generators, e.g., via tertiary control reserves or real-time mar-
Ref. [24] experimentally investigated the feasibility of of- kets. However, this is not always the case because (i) real-time
fering up and down regulation products with university campus markets might not exist and/or (ii) activation of tertiary control
buildings, and identified baseline prediction and latency as reserves might not be economical. In addition, as RES pene-
potential obstacles for implementation. The authors of [25], tration increases, the biases in RES forecast errors will intro-
[26] investigated SFC via fan power control in buildings duce more bias in the SFC signal. With fewer controllable gen-
with variable air volume (VAV) systems. Based on simula- erators in the production mix, other resources will have to ab-
tions, the authors found that up to 15% of a building's fan sorb these biases. To address this important problem, we pro-
power can be offered as reserves without significant loss of pose a method that allows building aggregations to systemati-
comfort, if the SFC signal is within the frequency band cally schedule their reserve capacity depending on how much
. In [27], a control approach similar bias exists in the SFC signal. The method can handle energy-
to that of [25], [26] was experimentally validated in a real constrained frequency signals, which is a significant advantage
building, but without considering frequencies below 1/(10 min) of our method compared with [30] that neglected energy con-
to avoid effects on chiller power consumption. Baseline esti- straints, and [31] that expressed the capacity as a function of
mation was performed in [27] using a low-pass filter, and the the AS signal's frequency.
experimental results showed that communication delays were The third contribution is the framework's application in a
not an obstacle for SFC provision for the considered building simulation study to estimate the SFC reserve potential by dif-
and frequency range. The follow-up work [28] included chiller ferent building types with different monetary incentives, and for
control, which enlarged the frequency band of SFC signals to different reserve product characteristics such as duration, sym-
1/(60 min). Ref. [29] investigated SFC by direct control of a metry, and energy content. The estimated amount of reserves is
heat pump's compressor power using a variable speed drive an upper bound because an idealized simulation setting is as-
motor in a water-based HVAC system. sumed where (i) there is no plant-model mismatch, and (ii) the
Apart from accurate tracking of the SFC signal, a TSO needs weather and occupancy predictions are perfect.
guarantees that the reserve capacity of commercial buildings This paper builds on our previous work [32] but includes sub-
will be reliably available. For this reason and since buildings stantial extensions. The main additions lie in (i) modeling the re-
are energy-constrained resources, a priori reserve scheduling is serve uncertainty, (ii) considering multiple actuators providing
required to maintain occupants' comfort. In this direction, [30] reserves in each building, and (iii) performing more detailed
developed an MPC-based method to quantify the flexibility of simulation studies. Although with appropriate modifications the
a commercial building and a contractual framework to declare framework can serve as a basis for implementation on a real
it to the utility. building, this paper does not focus on implementation issues.
In this paper, we follow this line of research and investigate The rest of the paper is organized as follows: Section II sum-
how SFC reserves can be provided by aggregations of com- marizes some important aspects of power system SFC,
mercial buildings. The problem we consider is how a building Sections III and IV introduce the modeling and control
VRETTOS et al.: ROBUST ENERGY-CONSTRAINED FREQUENCY RESERVES FROM AGGREGATIONS OF COMMERCIAL BUILDINGS 3
framework, the performance of the framework is demonstrated of reserve, especially for energy-constrained SFC signals with
and evaluated in Section V in a simulation example, a sensitivity limited reserve energy utilization.
analysis is performed in Section VI, and Section VII concludes
the paper. III. MODELING
II. FREQUENCY CONTROL IN POWER SYSTEMS
A. Motivation for Aggregations
Typically, a TSO controls frequency in three steps: primary,
secondary and tertiary control. Primary control is a distributed, We are interested in reserve provision from building aggrega-
proportional controller that stabilizes the frequency after a tions for two main reasons. First, the reserve markets typically
disturbance. Secondary control is a centralized controller that have requirements on the minimum size of the bid reserve ca-
restores the frequency to its nominal value and maintains the pacity. In most cases, individual commercial buildings cannot
desired exchanges between neighboring control areas. Tertiary meet these minimum bid size requirements, so building aggre-
control releases secondary control in case of large disturbances gations are needed to access the market. Even in markets with
and is typically manually activated. Before continuing, we low minimum bid size requirements, e.g., in the range of a few
summarize some important aspects of scheduling and activa- hundred kWs, building aggregations would still be of interest in
tion of SFC in power systems [33], which will be particularly presence of other typical requirements, such as symmetry and
relevant to the sensitivity analysis of Section VI. minimum duration of the bid reserve capacity. As shown in [32],
In most of Europe, SFC reserves are procured in a market aggregating buildings with different characteristics results in a
setting, i.e., the generators bid their reserve capacity and price in larger total reserve capacity compared with the case where each
weekly or daily auctions. The requirements of the auctions and building participates individually in the market.
SFC reserve products vary among countries. The minimum bid The second argument in favor of building aggregations is
size is typically in the range 1 to 10 MW, e.g., 5 MW in Switzer- more practical. An aggregator's job would be to determine the
land [34]. In many countries, only symmetric reserves, i.e., reserve capacity, bid it in the reserve market, and interact with
equal up- and down-reserve capacities, are allowed, whereas in the TSO during reserve activation and for the financial settle-
other countries asymmetric reserves are also accepted. ment. These tasks are very different to the normal activities of a
The reserve is requested from the generators via a signal sent building manager; therefore, the aggregator could take over this
by the TSO, typically every 2–4 seconds. There are two main burden that would otherwise be with the building manager.
activation rules: (a) the pro-rata activation, where the reserve
B. Building Aggregation Model
request is proportional to the capacity, and (b) the merit-order
activation, where the reserve is requested based on the short We consider buildings with integrated room automation
term marginal costs of generators. In this paper, we treat the (IRA), where heating, cooling, ventilation, blinds, and lighting
SFC signal as uncertain because it is unknown at the time when are jointly controlled. IRA is typically used in office buildings
the reserve capacities are scheduled. because it provides high comfort while being energy efficient
It is a standard practice today to have a single SFC signal in a [41]. We represent the thermal dynamics of a building with a
control area, which is typically the output of a proportional-in- 12th order multiple-input-multiple-output RC model [14].
tegral (PI) controller with the area control error (ACE) as input The model's state vector at time step is denoted by
[35]. However, a TSO could alternatively provide a number of , where is the number of states, and it includes the
reserve products and each provider could choose the product room air temperature as well as the temperatures in different
to offer its reserve capacity. This can be achieved by decom- layers in the walls, floor, and ceiling (all measured in ). The
posing the original signal into multiple SFC signals with dif- model's input vector includes the IRA control inputs,
ferent energy contents and ramping rates using filtering or opti- namely heating and cooling power, ventilation, blind position,
mization-based techniques [31], [36]–[39]. Such an approach is and lighting ( is the number of actuators). The heating and
already applied in the PJM control area, where the SFC signal cooling are represented in the thermal model as heat fluxes af-
is split into a slow (RegA) and a fast component (RegD) [40]. fecting the system states and their units are , i.e., the heat
Typically, the provision of SFC reserves is remunerated sep- fluxes are normalized by the floor area. The blind position is a
arately for the accepted reserve capacity in the auction (avail- number between 0 (fully closed) and 1 (fully open). The lighting
ability) and the requested reserve energy by the TSO (utiliza- is also normalized by the floor area and measured in . The
tion). In some countries the reserve capacity auctions are pay- model's disturbance vector includes the disturbances
as-bid, whereas in others the auctions have a common clearing that affect building states, e.g., ambient temperature in , solar
price. The remuneration of reserve energy is also country depen- radiation in , and internal heat gains by occupants and
dent; for example, in Switzerland the reserve energy remunera- equipment in ( is the number of disturbances). The
tion is coupled with the energy price in the spot market. In some output of the system is defined as and it includes the
cases, e.g., in PJM, the reserve provider is additionally remuner- room temperature in and illuminance in lux.
ated based on the SFC tracking performance. In this paper, we Because of thermal inertia, heating/cooling actuators can be
do not consider performance-based payments because they are used to provide frequency reserves. Denote by the
not typically used in the European market setting that we focus uncertain change in heat fluxes due to change in power con-
on. Furthermore, we neglect the reserve energy payments be- sumption of the heating/cooling devices during reserve provi-
cause the capacity payments are usually much higher per unit sion, where is the number of actuators that are used
for reserve provision. In buildings with water-based HVAC sys- the resulting convex linear program (LP) is solved until con-
tems, the fresh air flow rate for ventilation purposes is usually vergence, i.e., until the change in optimal solution between two
kept constant because changes would be immediately realized iterations is below a threshold [41].2
by the occupants. For this reason, we do not use ventilation for Denote by the predicted state of building
reserve provision. for time at time . The predicted states at time along
Denote by the (electric) reserve capacity of actuator of a prediction horizon are assembled in one vector as
building at time step . Buildings can provide up-reserves by . Adopting the same no-
decreasing their consumption, and down-reserves by increasing tation for inputs and disturbances, the building dynamics along
it.1 For now, we assume symmetric reserve capacities; asym- can be written as and
metric reserves will be discussed in Section IV. Since the HVAC , where the matrices , , ,
control input for heating and cooling is defined as a heat flux, it , , , and are of appropriate dimensions. The con-
is convenient to define also the “thermal” reserve capacity straints on outputs (thermal comfort zone) and HVAC control
that has units. can be obtained from by divi- inputs along are
sion with the coefficient of performance (COP). For notational
convenience, we use the variable in the problem formula- (3)
tions and call it simply reserve capacity keeping in mind that
where has 0 or 1 as entries. By substituting the
it is actually the “thermal” reserve capacity. In the pro-rata ac-
dynamics in (3) and defining
tivation case, the reserve energy is proportional to the reserve
capacity based on a normalized SFC signal . Note (4)
that the reserve capacity is a decision variable for the build-
(5)
ings, whereas the normalized SFC signal is uncertain. The
primitive uncertainty results in an uncertain change in elec- the HVAC and comfort zone constraints can be written in terms
tric power consumption of actuator , and the of the control inputs and the uncertainty as
corresponding uncertain change in heat flux .
With the above notation, the dynamics of a building are (6)
described by the linear time-varying model
where the matrices , , and are defined as
(1)
(2)
for appropriate matrices , , , , , and . The

(7)
matrix consists of the columns of that correspond to the
and is the N-dimensional identity matrix. Using , , and
heating/cooling actuators that provide reserves. The model (1),
, the HVAC input and comfort zone constraints along the pre-
(2) was validated against the well-known building simulation
diction horizon can be represented compactly in the optimiza-
software TRNSYS [14]. The model's satisfactory accuracy and
tion problems of Section IV.
its relatively low complexity make it suitable for MPC. A model
Consider an aggregation of buildings, i.e., .
similar to the one considered here was used in an MPC imple-
Denote by the vector con-
mentation in a real building, and it was found that it captures the
taining all predicted states of all buildings along . Using the
building thermal dynamics well [43]. Furthermore, this model is
same notation for inputs and disturbances, the input/output con-
flexibly customizable and allows us to perform large-scale sim-
straints of the aggregation can be written as ,
ulation studies with different representative building types.
where , , and are block diagonal matrices with , , and
Remark: The original model in [14] is bilinear between in-
on the diagonal, respectively. Denoting by
puts and disturbances (e.g., blind position and solar radi-
the SFC signal along , and by a collection of the
ation), as well as between inputs and states (e.g., blind posi-
reserve capacities of all actuators and all buildings along , the
tion and room temperature). For optimization purposes, a linear
input/output constraints of the aggregation can be written as
model is preferable, which is obtained applying a two-step pro-
cedure. First, we fix the disturbances to their predicted values
to eliminate the bilinearities between and , which results
in a time-varying model. To eliminate the remaining bilineari- ..
.
ties between and , we apply the so-called sequential linear
programming (SLP) approach. The bilinear terms are iteratively (8)
linearized around the most recently calculated trajectory and
2We use this SLP approach because [41], [43], [44] showed that the approxi-
1The(electric) reserve capacity is the amount of SFC reserves that becomes mation turns out to be very good since the non-linearities are mild. Note however
available to the TSO [42]. In the context of frequency regulation, the term up-re- that SLP is a heuristic approach and there are no guarantees for convergence to
serves denotes increase of a generator's production or decrease of a load's con- the globally optimal solution. In our simulations, we observed that convergence
sumption to increase system frequency. Similarly, the term down-reserves de- was usually achieved within 3 iterations. To speed up the simulations, we set the
notes decrease of a generator's production or increase of a load's consumption maximum number of iterations to 5, i.e., if the problem did not converge within
to decrease system frequency. 5 iterations we kept the best solution found so far.
where , is the reserve capacity D. Other Sources of Uncertainty

matrix, and is a column vector. The diagonal vec- Apart from the SFC signal, in general, there are two other
tors of are the actuators' reserve capacities for every time sources of uncertainty associated with our problem: (a) weather
step , i.e., . This lower triangular struc- and occupancy uncertainties, i.e., deviations from the predicted
ture satisfies causality and allows us to model the effect of past values, and (b) electricity and reserve price uncertainties. For
SFC signals. For example, accounts for the signal at time simplicity, weather/occupancy uncertainties are not considered
to determine the reserve at time . However, to comply with in this paper, i.e., the predictions are assumed to be perfect.
today's practice we fix for in the following, i.e., There exist several approaches in the literature of building cli-
the reserve energy depends only on the capacity and the current mate control to handle these uncertainties such as stochastic
SFC signal. Thus, is a block diagonal version of . MPC with additive Gaussian noise [41], scenario-based MPC
[45], or simply constraint tightening, which can be relatively
easily integrated in our framework.
C. Uncertain Frequency Signals
Following the common market structure, we assume that the
1) Power Constraints (PC): In pro-rata activation [34], the building aggregator acquires energy from the local utility com-
reserves are requested by the TSO via a normalized frequency pany but it sells its reserves directly to the TSO through the
signal . This box constraint represents the power wholesale reserve market. Since the utility company tariffs are
constraints of the signal. The values and typically constant over long periods, there is no electricity price
indicate full activation of up- and down-reserves, respectively. uncertainty in our case. However, the situation is different for
The uncertainty set along can be written as the reserve capacity prices. Although the reserve bid price is se-
lected by the aggregator, there is uncertainty involved because
(9) the bid might not be accepted in the auction, depending on the
offers of the other reserve providers. This uncertainty can be
2) Power and Energy Constraints (PEC): To facilitate the captured by using reserve price scenarios, e.g., based on his-
participation of energy-constrained resources (thermal loads torical data or via a market clearing model that uses as inputs
and storages) in AS, the frequency signals can be distinguished information about the offers of the other reserve providers. In-
by their energy content. A building aggregation could then corporating this source of uncertainty is beyond the scope of this
choose the preferred product to offer reserves. Energy con- paper, but it may well constitute an interesting topic for future
straints can be cast as linear constraints on the mean value of research.
the signal as
IV. CONTROL FRAMEWORK
(10) A. Hierarchical Scheme and Control Sequences

In this section, we present the hierarchical control scheme for
frequency reserve provision by commercial building aggrega-
where is the bias coefficient, is the averaging period, and tions, which is graphically shown in Fig. 1. Level 1 (Lv1), the
both are fixed by the TSO for a particular product.3 aggregator scheduling, is performed on a daily basis centrally
Equation (10) implies that the bias of the signal over is by an aggregator. There are two conflicting objectives in the
bounded. If we stack (10) along , we get the polyhedral con- problem: (a) minimize energy consumption through energy ef-
straint on the uncertainty , where is a matrix ficient control, and (b) deviate from the energy-optimal solution
with entries 1, 0 or 1, and is a vector with entries . The to leave slack for reserve provision. The optimal tradeoff be-
power constraints (9) are still present, since the full reserve ca- tween these conflicting objectives depends on the monetary in-
pacity could be requested anytime. Denote by the N-dimen- centives for reserves, which we call capacity payments, and the
sional identity matrix, and by the N-dimensional vector with cost of electricity. Lv1 determines the reserve capacity, and its
ones. Defining and , allocation among buildings, which achieve the optimal tradeoff
the uncertainty set for PEC is while respecting occupants' comfort in case the reserve is re-
quested. The aggregator solves a robust optimization problem
using building state measurements, prices of electricity and re-
(11) serves, as well as predictions of weather and occupancy. We
formulate the robust problem first for PC only and second for
Modeling the reserve uncertainty in the presence of energy PEC. Reliable provision of frequency reserves is critical for
constraints using the polyhedral set is a new contribution power system security, and so TSOs typically require 100%
compared with our previous work [32]. availability of the reserve capacity [34]. For this reason, HVAC
3This is a time domain approach to model the energy constraints of the SFC
input constraints are truly hard constraints and robust optimiza-
signal. An alternative approach in the frequency domain would be to multiply tion is a natural approach to follow.
the SFC signal with the transfer function of a high-pass filter. We adopt the Level 2 (Lv2), the building HVAC control, is a robust MPC
time domain modeling approach in the reserve scheduling and MPC problems [46] that determines the energy optimal HVAC control inputs
because it allows us to define the uncertainty set as a polyhedron (i.e., a set of
linear inequalities), which is in a standard form and so existing techniques from every 30 minutes locally at each building, and leaves enough
robust optimization can be directly used. slack for the reserves computed in level 1. Level 3 (Lv3), the
equivalent of (13) is . Repeating this

procedure for all rows of (12b), the robust counterpart problem
of (12) can be written as
(14a)
(14b)
(14c)
where and are defined by
Fig. 1. Overview of the proposed hierarchical control scheme. Signals and con- and . Problem (14) is an LP, but it grows
trol actions with dashed curves are executed in level 1 (i.e., once a day), the ones
with thin/solid curves are executed in level 2 (i.e., every 30 minutes), and the quadratically in since we need auxiliary variables and
ones with thick/solid curves are executed in level 3 (i.e., nearly in real-time, additional constraints for each uncertain constraint (13) to
typically every few seconds). model the 1-norm. A similar reformulation for a general class
of linear systems with reserve demands can be found in [48].
We now consider the case where either only heating or only
frequency signal filtering and tracking, is a feedback controller cooling actuators provide reserves in each building, which is
that tracks the SFC signal by controlling the power consumption likely in practice to avoid energy dumping by simultaneous
in real-time, e.g., every 10 seconds. In the remainder of this heating and cooling. Recall from (7), (4) that is a block di-
section, we present the control design to achieve the desired agonal collection of . If only
objectives in each of the three levels of the control hierarchy. heating (cooling) actuators are used for reserve provision, then
all entries of are non-negative (non-positive). Additionally,
B. Level 1: Aggregator Scheduling
all entries of and are non-negative by construction.
The aggregator's goal is to determine the optimal amount of Therefore, every row of contains either only non-negative
reserves to be offered in the market. Denote by or only non-positive entries and, by construction, the same
and the electricity cost and reserve capacity pay- holds for all entries of , i.e., or
ment vectors, respectively, where is the prediction horizon. . Based on the definitions of , ,
Note that efficiency factors incorporating actuators' COP and , and , and recalling that is non-negative, can be
building distribution system losses are included in and . equivalently written as the linear term
Lv1 can be cast as the robust LP
(12a)
(12b)
(12c)
(15)
where is the uncertainty set, i.e., . Equation
(12b) requires that HVAC input and comfort zone constraints where and denotes the element-wise absolute
are satisfied even in the worst case of uncertainty realization. value operator. In this case, the more general formulation (14)
By appropriately selecting in (12c), we can impose constant can be simplified to the following LP that has the same size as
reserve capacities over a period of time and/or the block diag- (12), and so can be solved efficiently.4
onal structure on discussed in Section III-B. The building dy-
namics in (12b) are decoupled among buildings; however, the
coupling comes via (12c) and the objective function. (16a)
Denote by the th row of any matrix or vector . We (16b)
derive the robust counterpart of (12) for PC and PEC. Consider (16c)
the th row of (12b) . The
term is a scalar and can be written as , where 2) Robust Counterpart for Power and Energy Constraints:
is a column-wise vectorized version of and With PEC, the uncertainty set is a polyhedron and duality
, where is the Kro- theory can be applied to derive the robust counterpart problem
necker product. Thus, the th row of (12b) is equivalent to [47], [49]. We write constraint (13) as an optimization problem
over
(13)
1) Robust Counterpart for Power Constraints: In the pres- (17a)

ence of PC only, the uncertainty set is given by (9), i.e.,
(17b)
is constrained in an -norm ball. In this case, we can maximize
the left-hand side of (13) analytically using the dual of -norm, 4Note that (16) is a special case of (14), and it is equivalent to formulation
i.e., the 1-norm [47]. Following this procedure, the deterministic (13)–(15) in our previous work [32].
By deriving the dual problem of (17) for all , we obtain the is defined similarly to but for a single building, and is a
following robust counterpart of (12) column-wise vectorized version of . Note that is fixed
(18a) in Lv2, and the right-hand side of the inequality is a constant.
2) Robust Counterpart for Power and Energy Constraints:
(18b) With PEC, the robust counterpart of (19) is
(18c)
(18d) (20a)
(18e) (20b)
where is the vector of dual variables. Problem (18) is an LP, (20c)
but its size grows quadratically in . Although this dualization (20d)
technique holds also for PC, we use the uncertainty set and
apply analytic maximization in that case since it results in fewer Problem (20) is similar to (18), but there are two main dif-
variables and constraints. ferences: first, is fixed; and second, and are
Remark: Asymmetric reserves can be modeled defining , time-varying. For a time step in the averaging interval
for down- and , for up-re- of length , (10) can be written as
serves. In case of PC only, the uncertainty set remains a polyhe- , where is known because the un-
dron and a tractable robust counterpart problem can be derived certainty up to is realized. Thus, the coupling constraint on
using analytic maximization. The reader is referred to [32] for depends on the energy content of the SFC signal
detailed formulations. In case of PEC, the non-linear constraint in the previous time steps of the averaging interval, and this de-
is needed to ensure that up- and down- reserves are pendence is accounted for in the MPC.
not requested simultaneously. Therefore, the uncertainty set is
not a polyhedron any more and the dualization technique cannot
D. Level 3: Frequency Signal Filtering and Tracking
be applied in this case to derive the robust counterpart problem.
In Lv3, the HVAC consumption is controlled around to
C. Level 2: Building HVAC Control provide reserves.We consider water-based HVAC systems that
The goal of Lv2 is to control the building in an energy ef- are common in Europe, but the reserve scheduling method ap-
ficient way, while ensuring that the scheduled reserves can be plies also to air-based systems [30]. The power consumption of
provided, if requested. Given the optimal reserve allocation that water circulation pumps in water-based HVAC systems is typ-
is fixed in Lv1, in Lv2 the HVAC control inputs are determined ically small, and so we directly control the heating or cooling
every 30 minutes by the robust MPC with prediction horizon devices, e.g., heat pumps (HPs), to provide reserves. The de-
(19a) sired electric power consumption of HP of building is
(19b) (21)
where and are the parts of and , respectively, that where is fixed from level 2, and
are relevant for building . The first input of the optimal con-
is fixed from level 1. In case of asym-
trol sequence of (19) determines the Lv2 setpoint of the HVAC
system for the next 30 minutes,5 . Problem (19) formu- metric reserves, is equal to the down-reserve capacity if
lates an MPC with open-loop predictions, i.e., the optimization , and equal to the up-reserve capacity if .
is performed explicitly over the control inputs . MPC with In a fast time scale, and depending on the HVAC system,
closed-loop predictions, i.e., optimization over affine policies the HP consumption can be controlled by modifying either the
of the uncertainty, showed minor or zero performance improve- water temperature setpoint at condenser's outlet or the refrig-
ment in this case, and so it is not used.6 In the following, we erant's flow rate via valves. We rely on the second approach that
derive the robust counterparts of (19) for PC and PEC. was experimentally shown to be able to track fast reference sig-
1) Robust Counterpart for Power Constraints: In this case, nals, e.g., SFC signals, in [52]. The reference HP electric power
the deterministic equivalent of (19) can be obtained by substi- can be tracked using a feedback PI controller with error
tuting (19b) with , where signal the difference between the reference power and the actual
electric power consumption of the HP.
5Although a consumption schedule is calculated in Lv1, the MPC of Lv2 In this paper, we use a causal Chebyshev filter to get the en-
can reduce the costs due to less uncertainty (recent reserve requests are known ergy-constrained signal , which is sent to the building aggre-
and better weather forecasts might be available) and, possibly, shorter optimiza- gation (see Fig. 1), but other filters might also be used. Note that
tion time steps. In case of plant-model mismatches, MPC additionally reduces
constraint violations due to its closed-loop operation. The MPC schedule is the although is a filtered signal, it does not mean that its bias
building's baseline consumption, and is communicated to the aggregator. Since
the baseline is a by-product of the formulation, baseline prediction methods that 6An alternative to the standard MPC with open-loop predictions is MPC with
have proven to be hard [24] are not required.An advantage of using MPC is closed-loop predictions, where the control input is parameterized as a nominal
that the baseline is determined ahead of time, in contrast to online baseline es- input plus a feedback term both of which are optimization variables. The feed-
timation approaches that use filtering [27]. Furthermore, the predictive nature back term is typically an affine function (or policy) of either the state or the
of MPC inherently accounts for rebound effects due to reserve provision when uncertain disturbance. The interested reader is referred to [50], [51] for more
calculating future HVAC setpoints [10]. information.
is zero but instead that its bias is bounded. The filter's transfer V. PERFORMANCE OF CONTROL FRAMEWORK
function is
A. Investigation Setup
(22)
We investigate the performance of the proposed control
where is the filter's order, and , are its coefficients that
framework in simulations with an aggregation of typical
depend on the pass-band edge frequency . From a TSO per-
Swiss office buildings. We consider two HVAC systems: in
spective, and can be chosen such that the resulting low-fre-
system A, heating is performed via radiators with coefficient
quency component (LF signal) and high-frequency component
of performance (COP) equal to 3, whereas cooling with cooled
(HF signal) have ramping rates and energy contents that are ap-
ceilings ; in system B, both heating and cooling
propriate for commercial buildings. In this paper, we fix
are performed using thermally activated building systems
since it showed good performance in preliminary simulations,
(TABS) with . We also differentiate between
and we consider different based on different averaging pe-
heavy (eh) or light (el) building envelope, high (wh) or low
riods in Section V-B.
(wl) window area fraction, and high (gh) or low (gl) internal
Apart from Lv3, the filtering is also taken into account in the
gains. In our simulations, we consider an aggregation of 6
higher levels of the control hierarchy. Recall that the filtered
large buildings (15000 each) with the following charac-
SFC signal is tracked every few seconds in Lv3, whereas the
teristics: , ,
decisions in Lv1 (reserve scheduling) and Lv2 (determination
, , ,
of optimal building setpoints) are taken every 30 minutes. Due
and . Typical occupancy profiles were
to this time scale separation, the important information from the
used, whereas weather data were provided by Meteoswiss (the
filtering of Lv3 that needs to be conveyed to Lv1 and Lv2 is only
Swiss federal office of meteorology and climatology). More
the integral of the SFC signal over this period, i.e., the bias of the
information regarding the buildings can be found in [32], [41].
signal. By formulating the energy constraint of the SFC signal
The temperature comfort zone during working hours is
as the linear inequalities (10), we can directly account for the
in winter and in summer, which is in
signal's bias in the optimization problems of Lv1 and Lv2 in a
accordance with the ASHRAE 55-2013 standard [55]. During
tractable way. Note that the bias coefficients for different aver-
non-working hours and weekends, the comfort zone is relaxed
aging periods in (10) can be empirically obtained by applying
to in both seasons. The optimizations are performed
the filter (22) on historical data of SFC signals using different
with a time step of 30 minutes, which is the discretization
.
step of building models (1), and the prediction horizons of
The merit of the robust reserve scheduling formulation is
Lv1 and Lv2 are fixed to (2 days) and (1
a guarantee that the solution will be feasible (i.e., no HVAC
day), respectively. We assume symmetric, daily reserves, i.e.,
input constraint or comfort zone violations will occur) for any
constant reserve capacity over a day, and capacity payments
reserve request that lives in the considered uncertainty set
10% higher than the electricity price, i.e., .
, if the reserve uncertainty is the only source of uncertainty
in the problem. However, this theoretical guarantee will not
hold in a practical implementation with additional uncertainties
coming from plant-model mismatches and imperfect weather/ B. Parameters of Energy Constraints
occupancy predictions. If a similar guarantee for a more real-
istic case is desired, then special care must be taken to build To apply energy constraints as in (10), we determine appro-
robustness also to these additional uncertainties. priate pairs of averaging period and bias coefficient based
The tracking quality of depends on HP's mechanical on the historical normalized SFC signals from the
delays, dead-times, ramping limits, and minimum down-times Swiss control area for 2009 and 2012. We consider six different
and/or run-times. If such HP dynamics are significant, tests sim- averaging periods, , and calculate
ilar to the ones in [52] can be performed to identify upper limits six sets of filter parameters in (22), one for each of the pass-band
on the frequency content of that result in good tracking. edge frequencies . For each value of , we filter the
In this case, the buildings can form a coalition with faster historical SFC signals using the corresponding filter (22) to get
resources, e.g., an aggregation of residential electric water the HF signals for 2009 and 2012. For each of the four signals
heaters (EWHs), exclude very high frequency components of and for each value of , we calculate as the largest absolute
the SFC signal with a band-pass filter, and send them to the mean value of the SFC or HF signal over any period , i.e.,
faster resources. The reader is referred to [53] for relevant . The results are summarized in
simulation results, and to [54] for a real-life demonstration of Table I. Notice that the original SFC signals can be significantly
SFC provision by EWHs. Since our focus is on robust reserve biased over periods of several hours. Note also that the biases
scheduling and MPC, we do not implement a PI controller in of the HF signals are significantly lower than those of the orig-
our simulations but assume that HP dynamics are negligible, inal SFC signals, and that the signals in 2012 are generally less
i.e., the reference of (21) and the SFC signal can be biased than in 2009. This is because from March 2012 the ACE
perfectly tracked if the comfort zone and HVAC constraints are of Switzerland is netted with that of other European countries
satisfied. before generating the SFC signal. We use the HF signal of 2012
TABLE I
BIAS COEFFICIENTS FOR THE SFC SIGNAL AND ITS HIGH-FREQUENCY
COMPONENT (HF) FOR DIFFERENT AVERAGING PERIODS
Fig. 4. HVAC heating power for building A3 during the winter week (the plot
starts on Monday).
Fig. 2. Left: a 4-hour extract of the original SFC signal, its low-frequency (LF),
and its high-frequency (HF) components for a sample day. Right: the moving
averages of the signals for the whole sample day. Fig. 5. Room temperature for building A3 during the winter week (the plot
starts on Monday).
the MPC, whereas the grey-dashed curves correspond to the

final values after tracking the HF signal. The grey envelope
shows the problem's robust region, which is defined as the re-
gion of the thermal comfort zone with the following property.
Fig. 3. Optimal reserve allocation among buildings during the winter week (the If the temperature is within the robust region at , then for
plot starts on Monday).
any SFC signal with the appropriate energy content there ex-
ists a feasible control input trajectory that satisfies the thermal
and select ( , ) for the simulations of this section, comfort and HVAC input constraints (the black-dashed curves
according to Table I. in both figures) along the prediction horizon . In
our simulations, the HVAC inputs and room temperature stay
C. Operation for Typical Weeks always within the robust region, which indicates that the robust-
We present simulation results for typical weeks in winter and ness to reserve uncertainty was built properly.
summer. In Fig. 2, we show a 4-hour extract of the SFC, LF, In summer, the reserve potential from cooling actuators is
and HF signals for a sample day, as well as the 2-hour moving also significant. During the considered week, the capacities
averages of the signals for the whole day. Although the original range from approximately 60 kW on Sunday to 420 kW on
SFC signal is mostly negative, most of the bias is absorbed by Saturday with an average weekly value of 323 kW. We ob-
the LF signal, and so the HF signal is significantly less biased. served a more uniform distribution of reserves among different
The bias of the SFC signal is larger than 0.3, whereas the buildings in summer, compared to winter, because cooled
energy-constrained HF signal has a bias less than 0.15. ceilings and TABS have similar time constants. Due to space
Fig. 3 shows the optimal total reserve capacity of the aggre- limitations, plots similar to Figs. 3, 4 and 5 are omitted.
gation and its allocation among buildings for the winter week. Based on our results, the considered 6 buildings with floor
The capacity is constant for each day and ranges from approx- area 15000 m , average rated heating power 27 W/m , and av-
imately 260 kW on Friday to 380 kW on Saturday with an average rated cooling power 32 W/m provide a reserve capacity
erage weekly value of 313 kW. Note that the capacity is shifted equal to 313 kW and 323 kW on average in winter and summer,
among buildings in a way that maximizes the total capacity respectively. Using simple linear extrapolation, we find that the
of the aggregation. Interestingly, the buildings offer higher re- required minimum reserve capacity of 5 MW in Switzerland can
serve capacities when they normally consume less power. For be provided by 96 (similar) buildings in winter and 93 build-
example, buildings with system A contribute mainly at night, ings in summer, i.e., approximately 100 buildings in both sea-
whereas buildings with system B offer more reserves during sons. Note that this is only a rough estimate as: (a) it is based
working hours, because they prefer to preheat at night. The on the average reserve capacity values; and (b) the extrapolation
scheduled total energy consumption of the aggregation is spread might overestimate the required number of buildings to meet the
throughout the whole week to maximize the reserve potential. 5 MW limit, because the larger the aggregation the more flexi-
During the weekend, the buildings are not occupied and the bility exists in allocating reserves among buildings.
comfort zone is larger, which results in higher reserve capac-
ities compared with working days. D. Effect on Energy Consumption
The operation of building A3 based on Lv2 and Lv3 is pre- To maximize the reserve capacity, the buildings operate in a
sented in Figs. 4 and 5. Fig. 4 shows the HVAC system heating less energy efficient way. For example, during the considered
power and Fig. 5 the temperature trajectories. The black-solid winter week reserve provision resulted in a consumption 60%
curves indicate the scheduled consumption and temperature by higher than that of energy efficient building control, i.e., an MPC
(a) (b)
Fig. 7. Bid curves of building aggregation in winter and summer. (a) For PC
(a) (b) and PEC, and ratios . (b) For PEC and ratios .
Fig. 6. (a) Dependence of increase in energy consumption on the reserve ca-
pacity offered by the buildings. (b) Effect of increase in energy consumption on E. Discussion on Imperfect Disturbance Predictions
the average room temperature.
As mentioned in Section IV-D, the robust formulation guar-
antees satisfaction of the HVAC and comfort constraints for any
with objective to minimize electricity cost without offering re-
admissible reserve request and perfect weather/occupancy pre-
serves. As shown in Fig. 4, in order to provide reserves in both
dictions. However, the framework cannot provide mathematical
directions, the buildings try to operate close to the middle of the
guarantees on the worst case temperature deviations in case of
heating/cooling device's power range. This is in contrast to an
imperfect predictions.
energy efficient operation, where the power consumption would
An empirical estimate of the probability of the deviation and
be as close as possible to the minimum value. However, the in-
its magnitude could be provided by Monte Carlo simulations:
crease in energy consumption does not mean that the building
(i) schedule the reserve capacities using imperfect disturbance
control is suboptimal. For the given electricity price and ca-
predictions; (ii) simulate the building operation under distur-
pacity payment, this building operation minimizes the total cost
bance uncertainty, i.e., the weather and/or occupancy realiza-
defined as the sum of electricity cost and reserve profit.
tions are different to the predictions; and (iii) analyze the re-
In fact, there is a tradeoff between increasing the reserve ca-
sults to keep track of the number and magnitude of temperature
pacity and reducing energy consumption, which depends on the
deviations. Alternatively, probabilistic guarantees could be ob-
capacity payment. To illustrate this, we report the increase in
tained by modeling the weather and/or occupancy uncertainty
energy consumption for different reserve capacity levels (ex-
via scenarios and then robustifying against the reserve uncer-
pressed as a percentage of the maximum possible capacity) in
tainty separately for each scenario. This is an exciting research
Fig. 6(a). Note that reserve capacities up to 15% of the max-
direction for future work.
imum capacity do not practically increase the energy consump-
However, some intuition can be provided without following
tion, that 50% of the maximum capacity can be offered with
either of the above two approaches. Consider a building in en-
approximately 30% increase in energy consumption, and that
ergy efficient operation using a deterministic MPC that relies
in order to maximize the reserve capacity the energy consump-
on an imperfect weather forecast. As explained in Section V-D,
tion must be increased by 60%. The effect of increasing energy
the building would operate close to the minimum power con-
consumption is a higher average room temperature during the
sumption, and thus the temperature trajectory would stay close
considered winter week, as shown in Fig. 6(b). In theory, the in-
to one of the comfort zone boundaries. Now consider the same
crease in energy consumption is proportional to the room tem-
building operated under reserve provision. As shown in Fig. 5,
perature increase with a proportionality constant equal to the
the building would operate closer to the middle of the comfort
building's thermal resistance. This linearity is reflected in the
zone to maximize the reserve capacity. For this reason, we ex-
data points of Fig. 6(b). The reason why the rightmost data point
pect the reserve provision case to result in smaller and perhaps
of Fig. 6(b) does not fall on the same line as the other points
less frequent comfort zone violations compared with the en-
is because in that case the blinds are set to different positions,
ergy efficient control case, if the building is exposed to the same
which affects the building's overall thermal resistance.
weather/occupancy uncertainty in both cases.
Note that the reported energy consumption increase is in com-
parison with the energy consumption of a building controlled by VI. SENSITIVITY ANALYSIS
an MPC minimizing energy costs. If the comparison was per-
formed considering supervisory rule-based controllers, which A. Capacity Payments
are used in most buildings today, a smaller increase in energy As explained in Section IV-A, the proposed reserve sched-
consumption would be expected. uling methods identify the optimal tradeoff between minimizing
In addition, note that the energy consumption increase re- energy consumption and leaving enough slack for reserve provi-
ported here is qualitatively different to the results of [56]. The sion. The buildings would not deviate from the energy efficient
authors of [56] calculated the round-trip efficiency of buildings control and would not offer any reserves if the additional elec-
while offering fast DR services, i.e., the efficiency loss due to tricity cost occurring due to this deviation were higher than the
the utilization of reserves. On the contrary, in this paper, the reward received for the slack provided as reserve capacity. In
reported increase in energy consumption is the result of sched- principle, the amount of reserves depends on the ratio between
uling the building consumption in a way that large amounts of the capacity payment , i.e., the remuneration for each kW of re-
reserves can be offered reliably. The energy consumption in- serve capacity provided, and the electricity price , i.e., the cost
crease we report will occur even if no reserve is actually called, for each kWh of electricity consumed. In this Section, we in-
i.e., it is the price to pay for the availability of reserves. vestigate this relationship by running simulations, similar to the
ones in Section V-C, over 2-week periods in winter and summer TABLE II
and for various ratios. The total reserve capacity for each CAPACITY (MW) OF RESERVE PRODUCTS (POWER CONSTRAINTS)
case, i.e., the sum of the capacities of each day of the 2-week
period, is presented in Fig. 7, where the left plot (Fig. 7(a)) is
for ratios and the right plot (Fig. 7(b)) is for ratios
. These plots represent the aggregation's bid curves be-
cause they communicate how much capacity the aggregation is
willing to bid in the reserve market depending on the payment
it receives for each kW of the capacity.
TABLE III
For the simulations are performed for winter and CAPACITY (MW) OF RESERVE PRODUCTS (POWER & ENERGY CONSTR.)
summer with and without consideration of energy-constrained
SFC signals: in Fig. 7(a) the black curves correspond to winter
weeks (“win”) and the grey curves to summer weeks (“sum”),
whereas the dashed curves are for PC and the solid curves for
PEC. Our simulations show that with the same financial incen-
tive and for both seasons, the buildings are willing to offer up to
10% more reserves compared with PC if energy constraints are
considered. Note that the gap between PC and PEC is generally
larger for lower , particularly in winter. In winter, the capacity results are shown for PC in Fig. 7(b) because the capacity is
saturates at its maximum value at , whereas in summer zero for , as explained before.
it increases monotonically as the ratio increases up to 2. Fig. 7(a) shows that for unconstrained SFC signals (PC) a
The analysis of Fig. 7(a) focused on , which is a ratio , which means a capacity payment 1% higher
necessary condition for reserve provision with PC. This obser- than the electricity price, taps most of the reserve potential. In
vation was also made in our previous work [32]; however, here particular, further increasing the ratio up to the maximum con-
we provide an explanation by studying the structure of problem sidered value increases the reserves only by 8% in
(14). If , the optimal solution is , winter and no more than 18% in summer. Assuming an average
where is the energy optimal scheduling. If , the op- electricity price of 146.6 CHF/MWh, which is the case for con-
timal cost is 0 and any solution within the feasible range of sumers who consume more than 60 MWh/year in Zurich, with
will be optimal. If , the optimal solution is , a ratio capacity payments around 148 CHF/MW/h
where is the upper limit of , and the optimal cost will be are needed. This is significantly lower than the most expensive
, i.e., the aggregation earns profit. The limit accepted bids, but approximately 4 times higher than the av-
depends on input/output constraints, and so different so- erage capacity payment in 2013 [57]. Fig. 7(b) shows that en-
lutions are obtained for different ratios, as shown in ergy-constrained SFC signals can reduce the necessary capacity
Fig. 7(a). In case of daily reserves, need not to be satis- payments down to 32% of the retail price , but of
fied point-wise throughout the whole day; instead, can course with a large reduction in the reserve capacity.
be chosen during daytime, and at night when electricity Our results show that reserves are actually costly for build-
prices might be lower. Reserve provision will be triggered if ings already equipped with MPC for energy efficient (optimal)
, i.e., the capacity payment is on average higher control, especially if the SFC signal is not energy-constrained.
than the electricity price. However, note that the calculations are based on the prevailing
On the other hand, with energy-constrained SFC signals case where the buildings acquire energy in the retail electricity
(PEC) reserves can be provided also with ratios market. In another market setting where the buildings acquired
throughout the whole day. It is easier to explain this with an ex- energy directly in the spot market, the buildings could offer re-
ample. Assume that the buildings have declared a capacity serves at more competitive prices because the retail electricity
for day . Assume also that up regulation (i.e., consumption de- prices are typically significantly higher than the wholesale spot
crease) is mainly requested during day . If the signal is energy electricity prices.
constrained, only a fraction of the worst case reserve energy The analysis of this section provides intuition on the rela-
will be requested as consumption decrease from tionship between the amount of reserves from building aggre-
the buildings. The remainder of will be stored as gations and the capacity payments. In practice, estimating the
thermal energy in the buildings and will reduce the required capacity payment is a challenging task that needs to consider
heating/cooling energy (and the respective costs) during day additional costs, e.g., due to device wear, but also the competi-
. For this reason, the buildings are willing to provide tion, i.e., the expected bid prices of generators and/or other load
reserves even if . We present simulation results for aggregations.
PEC in winter and summer in Fig. 7(b). The threshold ratio for
reserve provision is for both winter and summer, B. Reserve Product Characteristics
and it depends on and . The capacity increases slowly in In this section, we fix and investigate the influence
the ratio range [0.32–0.99], particularly in summer, and then of important reserve product characteristics on reserve capaci-
it suddenly jumps to higher values as approaches to 1. No ties. For PC, we consider reserves with daily or hourly duration,
As expected, increasing generally decreases the reserve ca-

pacity. The same can be obtained by different pairs:
for example, both (8 h, 0.25) and (4 h, 0.5) obtain ,
but the first achieves a capacity 96.72 MW, whereas the second
achieves 93.61 MW. In this case, the capacity is more sensitive
to than to . We compared all cases in winter and summer
with the same and found out that in 70% of them smaller
(a) (b)
is preferable to smaller . This means that buildings can cope
Fig. 8. (a) Sensitivity of reserve capacity (shown by the color map in MW) on easier with signals that are significantly energy constrained over
averaging period and bias in winter. (b) Capacity's dependence on in
winter and summer.
long periods than signals that are moderately constrained over
shorter periods.
VII. CONCLUDING REMARKS

and symmetric or asymmetric capacities. For PEC, we investi-
In this paper, we proposed a new framework based on ro-
gate reserves with daily or hourly duration, as well as different
bust optimization and robust MPC for scheduling and provision
pairs, based on Table I. We summarize simulation results
of secondary frequency control (SFC) reserves by the HVAC
with respect to the total reserve capacity for two weeks in winter
systems of aggregations of commercial buildings. The frame-
and summer in Tables II and III, where positive values denote
work incorporates tractable methods to account for energy-con-
up- and negative values down-reserves.
strained SFC signals, and relies on decentralized reserve pro-
For PC and symmetric products,7 hourly reserves increase ca-
vision to keep real-time communication requirements low and
pacities by 2.9% in winter and 3.3% in summer, compared with
preserve privacy.
daily reserves. In our simulations, we observed that if asym-
To show how the framework can be used to estimate the SFC
metric reserves are allowed, the aggregation provides signif-
reserve potential from commercial buildings, we considered a
icantly more down- than up-reserves in summer, whereas in
specific case study with typical Swiss office buildings and we
winter no up-reserves are provided at all.
run simulations during a winter and a summer week. In our
Compared with PC, SFC with PEC increases the reserves for
case study, significant amounts of SFC reserves could be re-
all pairs up to 3.8% in winter and 5.6% in summer. For
liably offered by approximately 100 buildings without loss of
given and , adopting hourly instead of daily reserves in-
occupant comfort. The case study results indicate that asym-
creases the capacities up to 2% in winter and 9.6% in summer.
metric reserves are preferable for buildings in case of traditional
Note that the increase is higher for large . Notice that shorter
unconstrained SFC signals. In addition, we observed that en-
are characterized by smaller products, and so constraint
ergy-constrained SFC signals can reduce the necessary capacity
(10) on the uncertainty becomes tighter. Therefore, one would
payments and increase the reserves significantly (for example,
intuitively expect that decreasing increases the reserve capac-
by up to 10% in our case study) compared with traditional SFC
ities monotonically. However, one has to consider that longer
signals. Furthermore, reducing the duration of reserve products
couple more optimization periods, which is why a monotonic
from 1 day to 1 hour also increases the reserves significantly (for
behavior is not observed in our simulation results. example, by up to 9% in our case study). Note that the reported
To gain deeper insight into the dependence of reserve capaci- numerical results are specific to the considered case study (e.g.,
ties on and , we ran simulations for two weeks in winter and the particular weeks) and the assumptions made, and so they are
summer with hours and varying from indicative.
0.05 to 0.5 with 0.05 increments, i.e., 60 combinations for each Our analysis was based on four main assumptions: (a) there
season. We assume daily, symmetric reserves and simulate only is no plant-model mismatch; (b) the predictions of weather and
Lv1, i.e., the reserve scheduling problem, whereas the building occupancy are perfect; (c) all building states can be measured;
HVAC control is not considered. The reason is that many of the and (d) the reaction of heating/cooling devices is fast and does
combinations of and are not achievable by the Chebyshev not cause any wear. Therefore, the reported results provide an
filter (22), and so Lv2 will likely be infeasible. However, our upper bound on the amount of reserves from buildings. For a
analysis provides intuition on the effect of and that could be real implementation, additional care must be taken for (a)–(d):
useful for filter design in a practical application. accounting for modeling and weather/occupancy prediction er-
In Fig. 8(a), we use a color map to show the total reserve rors, use of state estimators, and modeling of the fast dynamics
capacities for each of the simulated cases in winter. Similar re- of heating/cooling devices.
sults are obtained for summer, but are omitted here due to space The reserve scheduling problem considered capacity pay-
limitations. The dependence of reserve capacities on and ments, but neglected revenues from reserve energy utilization.
demonstrates a clear pattern: increasing any of the two parame- Incorporating the latter in the framework is possible if SFC
ters reduces the reserve capacity. Note that the capacity is more signal scenarios can be generated based on historical data,
sensitive to than to . To better illustrate this, we present and is expected to reduce the necessary capacity payments.
the results of Fig. 8(a) based on the product in Fig. 8(b). Additionally, comfort constraints can be relaxed as chance
7Note that the results of Table II are not directly comparable with the values
constraints allowing comfort zone violations with a small
reported in Tables 1 and 2 of [32], because different COP values have been used probability, which is typical for building climate control [41],
in the simulations of the two papers. [45], while keeping HVAC input constraints robust. Such
combination of robust and chance-constrained optimization is [22] Y. Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, and P. Haves,
likely to increase reserve capacities. We are currently working “Model predictive control for the operation of building cooling sys-
tems,” IEEE Trans. Control Syst. Technol., vol. 20, no. 3, pp. 797–803,
on inclusion of the above aspects in the framework, as well as Mar. 2012.
implementation and testing on a real building. [23] F. Oldewurtel, A. Ulbig, A. Parisio, G. Andersson, and M. Morari,
“Reducing peak electricity demand in building climate control
using real-time pricing and model predictive control,” in Proc.
IEEE Conf. Decision and Control, Atlanta, GA, USA, 2010, pp.
REFERENCES 1927–1932.
[24] S. Kiliccote, P. Price, M. Piette, G. Bell, S. Pierson, E. Koch, J. Carnam,
[1] D. Callaway and I. Hiskens, “Achieving controllability of electric H. Pedro, J. Hernandez, and A. Chiu, “Field testing of automated de-
loads,” Proc. IEEE, vol. 99, no. 1, pp. 184–199, Jan. 2011. mand response for integration of renewable resources in California's
[2] F. Schweppe, R. Tabors, J. Kirtley, Jr, H. Outhred, F. Pickel, and A. ancillary services market for regulation products,” Lawrence Berkeley
Cox, “Homeostatic utility control,” IEEE Trans. Power App. Syst., vol. Nat. Lab., Tech. Rep. LBNL-5556E, 2007.
PAS-99, no. 3, pp. 1151–1163, Mar. 1980. [25] H. Hao, T. Middelkoop, P. Barooah, and S. Meyn, “How demand re-
[3] Y. Makarov, C. Loutan, J. Ma, and P. De Mello, “Operational impacts sponse from commercial buildings will provide the regulation needs of
of wind generation on California power systems,” IEEE Trans. Power the grid,” in Proc. 50th Annu. Allerton Conf. Commun., Control, Com-
Syst., vol. 24, no. 2, pp. 1039–1050, May 2009. puting, Monticello, IL, USA, Oct. 2012, pp. 1908–1913.
[4] K. Kalsi, M. Elizondo, J. Fuller, S. Lu, and D. Chassin, “Development [26] H. Hao, A. Kowli, Y. Lin, P. Barooah, and S. Meyn, “Ancillary service
and validation of aggregated models for thermostatic controlled loads for the grid via control of commercial building HVAC systems,” in Proc.
with demand response,” in Proc. Hawaii Int. Conf. Syst. Sci., Wailea, Amer. Control Conf., Washington, DC, USA, Jun. 2013, pp. 467–472.
HI, USA, 2012, pp. 1959–1966. [27] Y. Lin, P. Barooah, S. Meyn, and T. Middelkoop, “Experimental
[5] E. Vrettos, S. Koch, and G. Andersson, “Load frequency control by evaluation of frequency regulation from commercial building HVAC
aggregations of thermally stratified electric water heaters,” in Proc. system,” IEEE Trans. Smart Grid, vol. 6, no. 2, pp. 776–783, 2015.
IEEE PES Innovative Smart Grid Technologies Europe (ISGT Europe), [28] Y. Lin, P. Barooah, and S. Meyn, “Low-frequency power grid an-
Berlin, Germany, Oct. 2012, pp. 1–8. cillary services from commercial building HVAC systems,” in Proc.
[6] C. Perfumo, E. Kofman, J. H. Braslavsky, and J. K. Ward, “Load man- IEEE SmartGridComm Conf., Vancouver, BC, Canada, Oct. 2013, pp.
agement: Model-based control of aggregate power for populations of 169–174.
thermostatically controlled loads,” Energy Conv. Manag., vol. 55, pp. [29] Y. Kim, J. L. Kirtley, Jr., and L. K. Norford, “Variable speed heat pump
36–48, 2012. design for frequency regulation through direct load control,” in Proc.
[7] J. Mathieu, S. Koch, and D. Callaway, “State estimation and control IEEE T&D Conf. Expo., Chicago, IL, USA, Apr. 2014, pp. 1–5.
of electric loads to manage real-time energy imbalance,” IEEE Trans. [30] M. Maasoumy, C. Rosenberg, A. Sangiovanni-Vincentelli, and D. S.
Power Syst., vol. 28, no. 1, pp. 430–440, Feb. 2013. Callaway, “Model predictive control approach to online computation
[8] E. Vrettos, J. L. Mathieu, and G. Andersson, “Control of thermostatic of demand-side flexibility of commercial buildings HVAC systems for
loads using moving horizon estimation of individual load states,” in supply following,” in Proc. Amer. Control Conf., Portland, OR, USA,
Proc. Power Syst. Computation Conf., Wroclaw, Poland, Aug. 2014, Jun. 2014, pp. 1082–1089.
pp. 1–7. [31] P. Barooah, A. Busic, and S. Meyn, “Spectral decomposition of de-
[9] F. Oldewurtel, T. Borsche, M. Bucher, P. Fortenbacher, T. Haring, M. mand-side flexibility for reliable ancillary services in a smart grid,” in
G. Vayá, J. L. Mathieu, O. Mégel, E. Vrettos, and G. Andersson, “A Proc. 48th Annu. Hawaii Int. Conf. Syst. Sci., Kauai, HI, USA, Jan.
framework for and assessment of demand response and energy storage 2015, pp. 2700–2709.
in power systems,” in Proc. IREP IX Bulk Power System Dynamics and [32] E. Vrettos, F. Oldewurtel, F. Zhu, and G. Andersson, “Robust provision
Control Symp., Rethymnon, Greece, Aug. 2013, pp. 1–24. of frequency reserves by office building aggregations,” in Proc. World
[10] M. Piette, J. Granderson, M. Wetter, and S. Kiliccote, “Responsive Congress Int. Fed. Autom. Control, Cape Town, South Africa, Aug.
and intelligent building information and control for low-energy and 2014, pp. 12068–12073.
optimized grid integration,” Lawrence Berkeley Nat. Lab., Tech. Rep. [33] “Survey on ancillary services procurement and electricity balancing
LBNL-5662E, 2012. market design,” [Online]. Available: www.entsoe.eu/publications/po-
[11] R. Kramer, J. van Schijndel, and H. Schellen, “Simplified thermal and sition-papers last accessed 23.01.2015
hygric building models: A literature review,” Frontiers of Architectural [34] “Basic principles of ancillary service products,” [Online]. Available:
Res., vol. 1, no. 4, pp. 318–325, 2012. www.swissgrid.ch last accessed 23.01.2015
[12] J. A. Crabb, N. Murdoch, and J. M. Penman, “A simplified thermal [35] P. Kundur, Power System Stability and Control, N. J. Balu and M. G.
response model,” Building Services Eng. Res. Technol., vol. 8, no. 1, Lauby, Eds. New York, NY, USA: McGraw-Hill, 1994.
pp. 13–19, 1987. [36] O. Leitermann and J. L. Kirtley, “Energy storage for use in load fre-
[13] M. M. Gouda, S. Danaher, and C. P. Underwood, “Building thermal quency control,” in Proc. Innovative Technol. an Efficient and Reliable
model reduction using nonlinear constrained optimization,” Building Electricity Supply, Waltham, MA, USA, Sep. 2010, pp. 292–296.
and Environment, vol. 37, no. 12, pp. 1255–1265, 2002. [37] H. Hindi, D. Green, and C. Laventall, “Coordinating regulation and
[14] B. Lehmann, D. Gyalistras, M. Gwerder, K. Wirth, and S. Carl, demand response in electric power grids using multirate model predic-
“Intermediate complexity model for model predictive control of tive control,” in Proc. Innovative Smart Grid Technol., Anaheim, CA,
integrated room automation,” Energy and Buildings, vol. 58, pp. USA, Jan. 2011, pp. 1–8.
250–262, 2013. [38] Y. Makarov, P. Du, M. Kintner-Meyer, C. Jin, and H. Illian, “Sizing en-
[15] G. Mustafaraj, J. Chen, and G. Lowry, “Development of room temper- ergy storage to accommodate high penetration of variable energy re-
ature and relative humidity linear parametric models for an open office sources,” IEEE Trans. Sustain. Energy, vol. 3, no. 1, pp. 1039–1050,
using BMS data,” Energy and Buildings, vol. 42, no. 3, pp. 348–356, 2009.
2010. [39] I. Avramiotis-Falireas, T. Haring, G. Andersson, and M. Zima, “Re-
[16] G. Goddard, J. Klose, and S. Backhaus, “Model development and iden- design of the automatic generation control scheme in the Swiss power
tification for fast demand response in commercial HVAC systems,” system,” in Proc. IEEE PES General Meeting, Washington DC, USA,
IEEE Trans. Smart Grid, vol. 5, no. 4, pp. 2084–2092, 2014. Jul. 2014, pp. 1–5.
[17] J. E. Braun, “Load control using building thermal mass,” J. Solar En- [40] “PJM Manual 11: Energy & Ancillary Services Market Opera-
ergy Eng., vol. 125, no. 3, pp. 292–301, 2003. tions,” [Online]. Available: www.pjm.com/~/media/documents/man-
[18] J. E. Braun, “Reducing energy costs and peak electrical demand uals/m11.ashx last accessed 23.01.2015
through optimal control of building thermal storage,” ASHRAE Trans., [41] F. Oldewurtel, “Stochastic model predictive control for energy efficient
vol. 96, no. 2, pp. 876–887, 1990. building climate control,” Ph.D. dissertation, Dept. Inf. Technol. Electr.
[19] F. B. Morris, J. E. Braun, and S. J. Treado, “Experimental and simulated Eng., ETH Zürich, Zurich, Switzerland, 2011.
performance of optimal control of building thermal storage,” ASHRAE [42] ENTSO-E, “Network code on load-frequency control and reserves,”
Trans., vol. 100, no. 1, pp. 402–414, 1994. [Online]. Available: https://www.entsoe.eu/major-projects/net-
[20] J. E. Braun, “A comparison of chiller-priority, storage-priority, optimal work-code-development/load-frequency-control-reserves/Pages/de-
control of an ice-storage system,” ASHRAE Trans., vol. 98, no. 1, pp. fault.aspx last accessed 07.06.2015
893–902, 1992. [43] D. Sturzenegger, D. Gyalistras, M. Gwerder, C. Sagerschnig, M.
[21] G. P. Henze, R. H. Dodier, and M. Krarti, “Development of a predictive Morari, and R. Smith, “Model predictive control of a Swiss office
optimal controller for thermal energy storage systems,” HVAC&R Res., building,” in Proc. REHVA World Congress Clima, Prague, Czech
vol. 3, no. 3, pp. 233–264, 1997. Republic, Jun. 2013, pp. 1–11.
[44] D. Gyalistras and M. Gwerder, “Use of weather and occupancy fore- Evangelos Vrettos (S'09) received the Dipl.-Ing de-
casts for optimal building climate control (Opticontrol): Two years gree in electrical and computer engineering from the
progress report,” Terrestrial Systems Ecology, ETH Zurich, Building National Technical University of Athens (NTUA),
Technologies Division, Siemens, Tech. Rep. LBNL-5662E, 2010. Greece, in 2010. After his graduation, he worked
[45] X. Zhang, G. Schildbach, D. Sturzenegger, and M. Morari, “Scenario- as a research assistant at the Electric Machines
based MPC for energy-efficient building climate control under weather and Power Electronics Laboratory of NTUA. He is
and occupancy uncertainty,” in Proc. IEEE Eur. Control Conf., Zurich, currently working towards the Ph.D. degree at the
Switzerland, Jul. 2013, pp. 1029–1034. Power Systems Laboratory, Swiss Federal Institute
[46] E. Camacho and C. Bordon, Model Predictive Control. Berlin, Ger- of Technology (ETH) Zürich, Switzerland, where
many: Springer, 2013. he joined in September 2011. He is interested in
[47] J. Löfberg, “Automatic robust convex programming,” Optim. Methods methods based on control theory and optimization to
and Software, vol. 27, no. 1, pp. 115–129, 2012. enable efficient demand response programs for integration of renewable energy
[48] X. Zhang, M. Kamgarpour, P. Goulart, and J. Lygeros, “Selling robust- sources in the power system.
ness margins: A framework for optimizing reserve capacities for linear
systems,” in Proc. IEEE Conf. Decision Control, Los Angeles, CA,
USA, Dec. 2014, pp. 6419–6424.
[49] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge Univ., 2004. Frauke Oldewurtel (M’10) was born in Oldenburg,
[50] M. Morari and A. Bemporad, Robust Model Predictive Control: Germany. She received the M.Sc. degree from
A Survey. Chapter in: Robustness in Identification and Con- the Georgia Institute of Technology, Atlanta, GA,
trol. London, U.K.: Springer, 1999. USA, in 2004, the Dipl.-Wirtsch.-Ing. (Electrical
[51] P. J. Goulart, E. C. Kerrigan, and J. M. Maciejowski, “Optimization Engineering and Business Administration) degree
over state feedback policies for robust control with constraints,” Auto- from the Technical University of Braunschweig,
matica, vol. 42, no. 4, pp. 523–533, 2006. Braunschweig, Germany, in 2006, and the Ph.D.
[52] H. Hagiwara, S. Kawachi, J. Baba, K. Furukawa, E. Shimoda, and S. degree from the Automatic Control Laboratory, ETH
Numata, “Fundamental study on controllability of heat pump power Zurich, Zurich, Switzerland.
consumption for fluctuated power compensation,” in Proc. IEEE Int. During her studies, she did internships with
Conf. Clean Electrical Power, Ischia, Italy, Jun. 2011, pp. 413–419. Siemens Transportation Systems, Braunschweig,
[53] E. Vrettos, C. Ziras, and G. Andersson, “Integrating large shares of Germany, and Siemens Corporate Technology, Munich, Germany. Afterwards,
heterogeneous thermal loads in power system frequency control,” in she was a PostDoctoral and Marie Curie Fellow with the Power Systems
Proc. IEEE Powertech Conf., Eindhoven, The Netherlands, Jun. 2015, Laboratory of ETH Zurich, Zurich, Switzerland, and at the Automatic Control
pp. 1–6. Laboratory, UC Berkeley, USA. Her research interests are in control of
[54] M. Geidl, “The Virtual Energy Storage Implemented by Swisscom En-
buildings and electricity grids.
ergy Solutions,” [Online]. Available: www.eeh.ee.ethz.ch/fileadmin /
user_upload / eeh/news_events/Kolloquien / 2014_10_21_Coll_ETHZ
_v4.0_PRINT.pdf last accessed 10.09.2015
[55] “CBE Thermal Comfort Tool,” [Online]. Available: http://comfort.cbe.
berkeley.edu/ last accessed 08.06.2015 Göran Andersson (M'86–SM'91–F'97) received the
[56] I. Biel, I. Hiskens, and S. Backhaus, “Round-trip efficiency of fast M.S. and Ph.D. degrees from the University of Lund,
demand response in a large commercial air conditioner,” Energy and Sweden, in 1975 and 1980, respectively.
Buildings, vol. 97, pp. 47–55, 2015. In 1980, he joined ASEA's, now ABB's, HVDC
[57] “Secondary Control Power (Weekly),” [Online]. Available: division in Ludvika, Sweden, and in 1986 he was
www.swissgrid.ch last accessed 23.01.2015 appointed full professor in electric power systems
at KTH (Royal Institute of Technology), Stockholm,
Sweden. Since 2000 he is full professor in electric
power systems at ETH Zürich (Swiss Federal Insti-
tute of Technology), where he also heads the power
system laboratory. His research interests include
power systems dynamics and control, power markets, and future energy
systems.
Dr. Andersson is a Fellow of the Royal Swedish Academy of Sciences and the
Royal Swedish Academy of Engineering Sciences. He was the recipient of the
IEEE PES Outstanding Power Educator Award 2007 and of the George Monte-
fiore International Award 2010.

Robust Energy-Constrained Frequency Reserves From Aggregations of Commercial Buildings

Uploaded by

Copyright:

Available Formats

Robust Energy-Constrained Frequency Reserves From Aggregations of Commercial Buildings

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Robust Energy-Constrained Frequency Reserves From Aggregations of Commercial Buildings

Uploaded by

Copyright:

Available Formats

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON POWER SYSTEMS 1

Robust Energy-Constrained Frequency Reserves

2 IEEE TRANSACTIONS ON POWER SYSTEMS

4 IEEE TRANSACTIONS ON POWER SYSTEMS

for appropriate matrices , , , , , and . The

where , is the reserve capacity D. Other Sources of Uncertainty

(10) A. Hierarchical Scheme and Control Sequences

6 IEEE TRANSACTIONS ON POWER SYSTEMS

equivalent of (13) is . Repeating this

1) Robust Counterpart for Power Constraints: In the pres- (17a)

8 IEEE TRANSACTIONS ON POWER SYSTEMS

the MPC, whereas the grey-dashed curves correspond to the

10 IEEE TRANSACTIONS ON POWER SYSTEMS

12 IEEE TRANSACTIONS ON POWER SYSTEMS

As expected, increasing generally decreases the reserve ca-

VII. CONCLUDING REMARKS

14 IEEE TRANSACTIONS ON POWER SYSTEMS

You might also like