Energy & Buildings 197 (2019) 99–111

Contents lists available at ScienceDirect

Energy & Buildings

journal homepage: www.elsevier.com/locate/enbuild

Development and evaluation of a generalized rule-based control

strategy for residential ice storage systems
Aaron Tam, Davide Ziviani, James E. Braun, Neera Jain∗
Ray W. Herrick Laboratories, School of Mechanical Engineering, Purdue University, 177 S. Russell Street, West Lafayette, IN 47907-2099, United States

a r t i c l e i n f o a b s t r a c t

Article history: In recent years, variable electricity pricing has become available to residential consumers to incentivize
Received 4 November 2018 demand reductions during midday peak hours. Thermal energy storage (TES) systems enable consumers
Revised 16 March 2019
to store cooling energy when demand is low and assist A/C operation during peak demand periods. How-
Accepted 18 May 2019
ever, the cost savings achievable using TES are highly dependent on how the system is operated for a
Available online 18 May 2019
given utility rate structure. This study investigates control strategies for a packaged chiller unit integrated
Keywords: with ice storage that leverage available residential utility rate structures in the U.S. to reduce consumer
Thermal energy storage electricity cost. The present work describes the development and evaluation of a generalized rule-based
Rule-based control control strategy inspired by the performance of an optimal controller that minimizes monthly electric-
Dynamic programming ity cost considering both time-of-use energy and demand charges. The generalized rule-based controller
Residential cooling systems is compared against the optimal controller as well as to heuristic control strategies for TES that were
originally developed for commercial buildings for a range of equipment cooling capacities, TES sizes, ge-
ographic locations, and residential utility rates. The total electricity cost is determined using a simulation
model that includes models for the chiller unit, ice storage tank, and secondary loop components, along
with a building load model. Results show that the generalized rule-based controller can approximate the
performance of the optimal controller within 20% for all cases tested, and within 10% of the optimal cost
in 53% of the cases tested. The controller also performs significantly better than the heuristic strategies
for commercial buildings that were evaluated.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction cost savings, cooling systems with integrated TES also offer other
benefits. High penetration of renewable electricity generation in
Thermal energy storage (TES) is a popular option for reduc- California and other states is leading to utility incentives that pro-
ing operating costs in commercial cooling systems; however, it is mote systems with better demand responsiveness, which can be
rarely considered in residential applications. One of the major rea- achieved with TES. The primary cooling equipment can also be
sons for this is the structure of utility rates. Commercial utility downsized due to the additional cooling capacity provided by the
rates typically have on-peak energy charges that incentivize off- TES, thereby providing an additional economic advantage for TES
peak energy use as well as demand charges that penalize high systems.
peak power consumption. By integrating TES with building chillers, The system under consideration is a residential-scale cooling
cooling (e.g. in the form of ice) can be generated and stored dur- system with integrated ice storage as depicted in Fig. 1. It consists
ing off-peak hours when electricity rates are low and then utilized of a packaged chiller, a water-glycol secondary loop, and an inter-
to provide cooling to the building during on-peak hours. Histori- nal melt ice-on-pipe storage tank. Points 1–4 in the dotted area
cally, residential utility rates have had a flat structure, so the in- represent the packaged chiller and points 5–8 represent the sec-
clusion of TES with residential cooling systems has not enabled ondary loop. The secondary loop connects the outdoor packaged
significant operating cost savings. However, in recent years, vari- chiller to the ice storage tank and an indoor air handling unit. A
able rates have become more available to residential customers, three-way valve at point 6 is used to control the charging and dis-
thereby changing the potential for cost savings. Beyond operational charging rate of the storage tank.
Many different control strategies have been developed for man-
aging the charging and discharging of TES systems in commer-
cial buildings as summarized in [17] and [21]. A particular em-
∗
Corresponding author. phasis has been on the use of optimal control strategies, including
E-mail address: neerajain@purdue.edu (N. Jain). model predictive control [5,13,22]. More recently, researchers have

https://doi.org/10.1016/j.enbuild.2019.05.040
0378-7788/© 2019 Elsevier B.V. All rights reserved.
100 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

Nomenclature

c Specific heat (kJ/kg-°C)

D Target demand cost ($)
ġ Internal gain (kW)
g Acceleration due to gravity (m/s2 )
h Delivery head (m)
J Integrated utility costs ($)
k Stage (-)
L Latent capacity (kWh)
m˙ Mass flow rate (kg/s)
N Stages in dynamic programming (-)
P Power consumption (kW)
Q˙ Heat transfer rate (kW)
Q Capacity (kWh)
Rd Utility demand charge ($/kW) Fig. 1. Schematic of the proposed cooling system with secondary loop and ice stor-
Re Utility energy charge ($/kWh) age.
T Temperature (°C)
t Time step (h)
UA Heat transfer coefficient (kW/°C) system, but also with available energy sources (photovoltaics) and
x State of charge (-) sinks (plug-in electric vehicles) [15]. However, while optimal con-
ε Heat transfer effectiveness (-) trol strategies can offer many benefits, they are typically computa-
γ Demand charge binary switch (-) tionally complex and difficult to implement. To that end, heuristic
η Efficiency (-) (rule-based) control strategies have also been developed and eval-
uated for commercial buildings that offer the benefit that they are
Subscript
more practical from an implementation perspective ([4,6–10,12], p.;
actual Actual conditions
[16]). These simpler heuristic strategies prioritize use of either the
amb Ambient
chiller or the storage to meet cooling loads during on-peak periods.
any Anytime
When using chiller-priority control, the operating cost savings are
C Charge mode
primarily due to a reduction in on-peak demand and energy use
charge Storage charging requirement
that result from downsizing of the chiller through the use of stor-
chiller Chiller
age. On the other hand, strategies that tend to prioritize the use
chiller,max Maximum chiller cooling
of storage generally have greater cost savings opportunities, but
chiller,mid Mid-peak chiller cooling
require load-forecasting to predict the amount of cooling needed
chiller,off Off-peak chiller cooling
during the peak hours. One of the more effective heuristic control
chiller,on On-peak chiller cooling
strategies for TES in commercial buildings is a hybrid strategy that
chw Chilled water supply
alternates between the chiller-priority and storage-priority strate-
cc Indoor cooling coil
gies based on a set of rules and operating conditions [6]. This strat-
cc,in Indoor cooling coil inlet
egy was developed and evaluated for commercial buildings but has
cc,out Indoor cooling coil outlet
never been evaluated for residential applications.
D Discharge mode
Two important differences between residential and commercial
d Demand charge
building applications that impact the performance of existing con-
e Energy charge
trol strategies are the schedule and intensity of power consumption
f Secondary fluid
apart from heating, ventilation, and air conditioning (HVAC). Com-
f,o Secondary fluid at storage outlet
mercial buildings have very high daytime occupant and non-HVAC
HVAC Heating, ventilation, and air conditioning
related power density (lights, computers, and other plug loads)
house House
that mostly disappear outside of normal working hours; therefore
i Inlet
peak power demand usually occurs within those working hours.
k Time index
On the other hand, residential buildings have much lower over-
load Cooling load
all occupant and non-HVAC power density that are more evenly
max Maximum
distributed over time. For commercial buildings, the relatively low
mid Mid-peak period
unoccupied non-HVAC power consumption means that charging
off Off-peak period
the storage at high rates does not result in peak power demand.
on On-peak period
However, applying existing control strategies (that employ simple
pump Pump power
charging strategies) to residential buildings can lead to high de-
pump,max Maximum pump power
mand costs when there are anytime or off-peak residential de-
rated Rated conditions
mand charges. Therefore, the objective of this research is to de-
s Phase change temperature
sign a heuristic control strategy for cooling systems with integrated
sc Secondary loop
TES that can minimize the total utility costs under any residen-
storage Ice storage capacity
tial utility rate structure. We do this by first designing an opti-
tank Storage tank
mal controller, based on dynamic programming, as a reference and
zone Indoor set-point
then extending the rule-based control strategy of Drees and Braun
[6] to explicitly handle anytime and off-peak demand charges. Us-
considered more complex building scenarios in which coordina- ing a simulated system model, we evaluate the proposed algo-
tion is required between not only the TES and the building cooling rithm’s performance relative to other heuristic control strategies
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 101

Fig. 2. Example of chiller operation using full storage control. Fig. 3. Example of chiller operation using chiller-priority control.

as well as the optimal controller for different geographic locations capacity relative to a conventional system. An example of chiller
and utility rates. operation using chiller-priority is shown in Fig. 3.
An outline of the paper is as follows. In Section 2, we review Load-limiting storage-priority control is an alternative and rel-
existing heuristic control strategies and present sample residential atively simple strategy that is applied to partial-storage systems.
utility rates acquired from a survey of available rates. The system It uses load-forecasting to maximize the use of storage while at-
model is described in Section 3. An optimal controller is presented tempting to minimize peak chiller cooling rates during the on-peak
in Section 4, and the generalized rule-based controller is presented period. The control strategy separates each day into an off-peak
in Section 5. We compare the different control strategies through a unoccupied period, off-peak occupied period, and on-peak occu-
set of simulated case studies in Section 6 and summarize conclu- pied period. The on-peak and off-peak periods are defined by the
sions in Section 7. utility rate structure, and the unoccupied and occupied periods re-
flect the occupancy of the building. Storage is charged during the
off-peak unoccupied period using the entire chiller capacity, is dis-
2. Background
charged during the off-peak occupied period using chiller priority
to minimize the use of storage, and is discharged during the on-
2.1. Existing heuristic control strategies
peak occupied period with the chiller operating at the minimum
possible constant rate necessary for the storage to be at its min-
Here we review existing heuristic control strategies that will be
imum charge by the end of the on-peak period. This strategy re-
incorporated into the generalized rule-based controller described
quires forecasting of the on-peak cooling loads and uses this fore-
in Section 4. These include full storage, chiller-priority, and load-
cast along with the available storage capacity to determine the on-
limiting storage-priority control. A full storage control strategy shifts
peak period chiller cooling rate as
all on-peak cooling loads to the off-peak period. This control strat-
ton
egy requires the system to be sized such that the storage can meet j=1
Q˙ on, j t − Qstorage
all on-peak cooling loads on the design day. The chiller operates at Q˙ chil l er,on = , (1)
ton
maximum capacity during the off-peak period on the design day to
fully charge the storage. This approach results in a much larger sys- where Q˙ chil l er,on is the rate of cooling from the chiller during the
tem capacity when compared to a partial storage system, in which on-peak period (kW), Q˙ on is the cooling load during the on-peak
the storage is sized to meet only a portion of the on-peak cool- period (kW), t is the time interval (h), Qstorage is the available
ing load. Full storage control is able to achieve the most operating storage capacity (kWh), and ton is the duration of the on-peak pe-
cost savings because the chiller is not operated at all during the riod (h). The chiller cooling rate can be updated over time dur-
on-peak period, but this comes at the expense of a higher initial ing the on-peak period with improved forecasts and changes in
system cost. An example of full storage control is shown below in available storage capacity. Compared to chiller-priority, this control
Fig. 2 in which the chiller provides no cooling capacity during the strategy maximizes the use of storage, and tries to discharge the
on-peak period (shaded in gray). storage completely every day. This control strategy also minimizes
A typical chiller-priority control strategy has two distinct oper- the on-peak demand charge of each day because it operates the
ating modes: charging and discharging. The charging mode begins chiller at a minimum constant rate. An example of chiller opera-
during the off-peak period when the building is usually unoccu- tion using load-limiting storage-priority control is shown below in
pied. The chiller then operates at full cooling capacity until the Fig. 4.
storage is fully charged. In discharging mode, the chiller is oper- Drees and Braun [6] presented a rule-based controller that
ated to meet the building cooling load. If the load exceeds the combines elements of the chiller-priority and storage-priority
chiller capacity, then storage discharge occurs at a rate sufficient strategies, along with a demand-limiting algorithm. The demand-
to meet the load. Although chiller-priority is far from optimal, it limiting algorithm requires a measurement of the total building
is simple to implement and does not require load forecasting. The electrical use. A simpler version of this strategy that does not re-
primary operating cost savings are achieved by reductions in on- quire the power measurement is described in the ASHRAE Hand-
peak energy use and peak demand that occur for days with high book of HVAC Applications [18] and is a starting point for the
cooling loads. These savings are possible in comparison to a con- strategy presented in the current paper. The rule-based controller
ventional system because the chiller and storage are sized such makes decisions based on the utility rate structure and the build-
that the chiller must operate nearly continuously at maximum ca- ing’s occupancy period. It uses the chiller-priority charging strategy
pacity for the design day. This leads to significant downsizing in during the unoccupied period, but employs different discharging
102 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

are the duration of the price periods and the difference in price
between different periods. There are four different types of TOU
rates: flat energy with TOU demand, TOU energy only, TOU energy
with anytime demand, and TOU energy with TOU demand.
We surveyed the different rate structures currently available to
residential customers across the United States using a database
(OpenEI.org) maintained by the National Renewable Energy Lab-
oratory [14]. Based on information from OpenEI.org, there were
only 25 states that had some form of variable rates for residen-
tial buildings in 2012. This number has increased significantly in
recent years, and variable rates are now available in all 50 states,
with more companies incorporating demand charges as well. All
variable utility rates are optional for customers. We identified six
different structures based on combinations of energy and demand
charges: 1) flat energy only with no demand (50 states); 2) flat
energy with flat demand (19: AK, AZ, CO, FL, ID, IL, IN, IA, KY, MN,
Fig. 4. Example of chiller operation using load-limiting storage-priority control.
ND, OH, OK, SC, SD, TX, VT, WA, WY); 3) flat energy with on-peak
and off-peak demand (3: CO, FL, NC); 4) time-of-use (TOU) energy
only (48 states); 5) TOU energy with flat demand (4: AL, CO, GA,
strategies during the occupied period depending on economic con-
SC); and 6) TOU energy with on-peak and off-peak demand (4: AZ,
siderations. The strategy minimizes the use of storage (i.e., chiller-
NC, VA, WI). In this paper, we will consider all six of the afore-
priority) if it is not cost effective to replenish the discharged en-
mentioned rates in evaluating the proposed generalized rule-based
ergy (e.g., during the off-peak period) and if the current chiller
control strategy.
load is less than a target load limit. Alternatively, if the use of stor-
age would lower daily energy costs, and there is sufficient stor-
age to meet the remainder of the load for the occupied period 3. System model
without operating the chillers, then the chillers are turned off and
the load is met by storage alone (full-storage control strategy). In this section, we describe the system model used for evalu-
However, if there isn’t sufficient storage to meet the remainder of ating the performance of different control strategies. It consists of
the integrated loads over the discharge period, then the controller component models for an ice storage tank, a packaged chiller, the
switches to a load-limiting control strategy where the chiller op- building thermal load, and a secondary loop.
erates at the minimum constant load necessary to fully discharge
the storage over the discharge period. Finally, if the use of stor- 3.1. Ice storage model
age is not economical but the chiller load will be greater than a
target limit, the chiller load is restricted to the maximum of the An internal melt, ice-on-pipe storage model was developed for
target limit and the load-limiting value necessary to avoid prema- the secondary loop following the approach described by West and
ture storage discharge. The target load limit is reset to zero for Braun [19] in which heat transfer effectiveness is used to calculate
each new billing period (i.e., month) and then reset over time as the limit on the rate of change of energy in the ice storage tank.
the maximum of the previous target and the current load. Com- The model only considers latent charging and discharging. The rate
pared to the simpler chiller-priority control strategy, the rule-based at which energy is removed from the storage tank is calculated by
controller requires forecasting of future loads but results in signif-
icantly greater savings. Q˙ tank = ε m˙ tank c f (Ts − Tchw ), (2)

where Q˙ tank is the storage charging (+) or discharging (-) rate

2.2. Residential utility rates (kW), ε is the heat transfer effectiveness (-), m˙ tank is the secondary
fluid flow rate through the storage tank (kg/s), cf is the secondary
Utility rate structures play a significant role in how cooling sys- fluid specific heat (kJ/kg-°C), Ts is the phase change temperature
tems with integrated storage are controlled. The most common (°C), and Tchw is the chilled water supply temperature (°C). The
structure for the residential sector is a flat rate, in which the price limit of the rate at which energy can be removed or added to the
is the same throughout the day. This rate can include a seasonal storage tank is reached when m˙ tank equals the total flow rate in the
or monthly change in price. A tiered structure is common for flat secondary loop. At each time step, the state of charge is calculated
rates in which the price is based on the amount of energy used by
in the billing cycle, and the price changes after a certain amount
of energy is consumed. Flat rates can be separated into flat energy Q˙ tank,k t
xk = xk−1 + , (3)
only and flat energy with anytime demand. For energy only, the en- L
ergy charge ($/kWh) is the only charge applied and is the typical
where xk is the state of charge (-), k is the time index, Q˙ tank,k is
rate for residential buildings. For flat energy with anytime demand,
the storage charging (+) or discharging (-) rate at time k (kW), t
a fixed demand charge ($/kW) is included in addition to the en-
is the time step (h), and L is the maximum change in internal en-
ergy charge, which is an additional cost based on the highest av-
ergy that can occur during charge or discharge (kWh). The outlet
erage power consumed over a 15-minute window throughout the
temperature of the tank at time k is then calculated by
month.
 
Many utilities also offer time-of-use (TOU) programs to residen- Q˙ tank
tial customers. These rates vary the price of electricity based on T f,o = Tchw + . (4)
m˙ tank c f
the time of day, usually separating the day into on-peak, off-peak,
and in some cases mid-peak periods. This structure encourages cus- West and Braun [19] developed correlations between the state
tomers to shift their power consumption to off-peak periods when of charge and the heat transfer effectiveness by curve-fitting poly-
the price of electricity is lower. The main features of TOU pricing nomial functions to test data of charging and discharging cycles
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 103

chiller’s maximum capacity and coefficient of performance (COP)

at any given ambient temperature and chilled water supply tem-
perature as shown in Eqs. (7)-(8).

Q˙ chil l er,max
= 6.82 − (1.43 × 10−2 )Tchw + (5.38 × 10−5 )Tchw
2
Q˙ rated
− (3.61 × 10−3 )Tamb − (1.11 × 10−5 )Tamb
2
(7)

COPactual
= 1.73 + (2.43 × 10−2 )Tchw − (7.01 × 10−5 )Tchw
2
COPrated
− (2.01 × 10−2 )Tamb + (2.49 × 10−5 )Tamb
2
(8)
The variable Q˙ chil l er,max is the chiller maximum capacity (kW),
Q˙ rated is the chiller’s rated capacity (kW), Tchw is the chilled wa-
ter supply temperature (˚F), Tamb is the ambient temperature (˚F),
Fig. 5. Ice storage charging and discharging heat transfer effectiveness curves. COPactual is the chiller’s coefficient of performance at the specified
operating conditions (-), and COPrated is the chiller’s rated coeffi-
cient of performance (-). The effect of chiller part-load conditions
given by on the COP is neglected in this model. The map is normalized so
that different chiller sizes and efficiencies can be easily modeled.
εC = 0.92 − 0.62x + 4.93x2 − 17.05x3 + 24.02x4 − 12.12x5 , (5) The chiller model considered in this study is based on a rated ca-
pacity of 3-tons (10.55 kW) and a COP of 3 with a rating condition
εD = 0.49 + 0.81x − 0.98x2 + 0.67x3 . (6) of 95°F (35 °C) ambient temperature and 45°F (7 °C) chilled water
supply temperature.
The subscript C represents charging and D represents discharg-
Fig. 6 shows cooling capacity and COP as a function of ambi-
ing. As shown in Fig. 5, there is a significant decrease in heat trans-
ent and chilled water temperatures. As the chilled water supply
fer effectiveness as the storage tank reaches full charge due to a
temperature decreases, the system loses capacity and efficiency be-
loss in surface area of the ice and water interface caused by the
cause of a lower evaporating temperature. Similarly, efficiency and
intersecting ice formations. Because of this decrease in heat trans-
capacity decrease with increasing ambient temperature due to a
fer effectiveness, the ice storage tank is oversized in this study so
higher condenser temperature. In short, there is an energy penalty
that 80% of the TES can be used to meet design day loads. There
for charging the storage tank that must be considered when evaluat-
is also a decrease in effectiveness in discharging mode as the TES
ing the performance of cooling systems integrated with TES.
is depleted, but the effect is smaller than for charging. Here we
only consider the latent capacity of the storage tank, so a state of
charge of 0 denotes a full tank of water at 0 °C (32°F), and a state 3.3. Building load model
of charge of 1 denotes a full tank of ice at 0 °C (32°F).
A simple conductance model with an internal gain was utilized
3.2. Packaged chiller model to generate building load profiles based on ambient temperatures.
Weather data for the typical meteorological year (TMY) was ob-
A residential-scale fixed-speed packaged chiller unit is modeled tained from the data set TMY3 provided by the National Solar Ra-
using the open-source ACHP (Air Conditioning/Heat Pump) tool diation Data Base [20]. Building load profiles were then calculated
[1]. ACHP uses detailed mechanistic models and includes thermo- according to
physical properties of working fluids from the CoolProp and Ref-
Q˙ load = U Ahouse (Tamb − Tzone ) + g˙ , (9)
Prop libraries [2,11] to calculate chiller performance and capacity
at different operating conditions. A performance map for the pack- where Q˙ load is the hourly cooling load (kW), UAhouse is the heat
aged unit considered here was developed using linear regression transfer coefficient (kW/ °C) calculated by choosing a design point
on the data generated from the ACHP. The map determines the of 3 tons (10.55 kW) at an ambient temperature of 35 °C, Tamb is

Fig. 6. Capacity and COP of the modeled packaged chiller at different ambient conditions and chilled water supply temperatures.
104 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

the ambient temperature (°C), Tzone is the indoor temperature set- rate of the pump, ηpump is the pump efficiency (-), g is the accel-
point (°C), and ġ is the internal gain of the house (kW), which is eration due to gravity (m/s2 ), and hmax is the maximum delivery
assumed to be a constant 1 kW. head (m).
The ice storage, chiller, and building load models are integrated
with the secondary loop model to form a complete system model.
3.4. Secondary loop and integrated system models
For a given building load, chiller cooling rate or tank charging rate,
and state of charge, the pump flow rate and indoor coil return tem-
The secondary loop employs a mixture of 70% water and 30%
perature are determined by iterative solution of the secondary loop
glycol as the working fluid. Energy balances are applied to the in-
energy balance equations Eqs. (10)–((13)), subject to the constraint
door coil and chiller evaporator to relate the heat transfer rates to
defined in Eq. (14) as well as those for the component performance
temperature changes across the secondary fluid according to
models (e.g. maximum chiller capacity, heat effectiveness of the ice
Q˙ load = m˙ sc c f (Tcc,out − Tcc,in ), (10) storage tank, etc.). The building load varies with time as a result
of ambient temperature variations, with hourly temperatures ob-
tained from TMY3 data [20]. The ambient temperature also influ-
Q˙ chil l er = m˙ sc c f (Tcc,out − Tchw ), (11)
ences the COP of chiller. In order to simulate the model, it is nec-
where Q˙ load is the hourly cooling load (kW), Q˙ chil l er is the chiller essary to specify an initial state of the storage tank and a control
cooling rate (kW), m˙sc is the mass flow rate in the secondary loop strategy for varying the chiller cooling rate over time in response
(kg/s), Tcc,in is the supply temperature to the indoor coil ( °C), and to varying building loads and utility rates. For each time step, the
Tcc,out is the return temperature from the indoor coil ( °C). system of equations are solved and then the state of charge is up-
The storage charging (-) or discharging (+) rate, Q˙ tank , is re- dated according to Eq. (3). The system power consumption for each
lated to the load and chiller cooling rates using an overall en- time step in the simulation period is calculated using Eqs. (15)–
ergy balance on the secondary loop according to Eq. (12). The stor- (18). A solution flow diagram which shows an inner loop solution
age charging and discharging rates are limited by the heat trans- scheme for the secondary loop energy balance and an outer loop
fer effectiveness model presented in Eq. (2) and change over time for each step in time is shown in Fig. 7.
with the state of charge. These rates also depend on the flow rate
through the storage tank, which is controlled to maintain a con- 4. Optimal control benchmark
stant temperature entering the indoor coil (state point 7 in Fig. 1)
as shown in Eq. (13). The mass flow rate through the storage tank An optimal controller was designed as a benchmark to de-
is constrained to be less than or equal to the secondary loop flow termine the minimum utility costs for providing cooling over a
rate as shown in Eq. (14). The chilled water supply temperature monthly billing period assuming that there are time-of-use en-
provided by the chiller is 25°F (−4 °C) for storage charging and ergy charges along with on-peak, off-peak, and anytime demand
45°F (7 °C) for storage discharging. The water glycol temperature charges. The problem is mathematically stated as
supplied to the indoor coils is set to be 5°F above the chilled wa- N
min J = Don + Do f f + Dany + k=1 Re,k PHVAC,k t
ter supply temperature to the storage tank during charging, and Don ,Do f f ,Dany ,
Q˙ tank,k for k=1,...,N (19)
5°F below the chilled water supply temperature to the storage tank
during discharging. s.t.

Q˙ tank,min,k ≤ Q˙ tank,k ≤ Q˙ tank,max,k (20)

Q˙ tank = Q˙ load − Q˙ chil l er (12)

0 ≤ Q˙ chil l er,k ≤ Q˙ chil l er,max,k (21)

m˙ sc Tcc,in = (m˙ sc − m˙ tank )Tchw + m˙ tank T f,o (13)

xmin ≤ xk ≤ xmax (22)

m˙ tank ≤ m˙ sc (14)

The total power consumption of the system is the sum of the x0 = xN (23)
power consumption of the chiller and the pump. The chiller power
consumption is calculated using the performance map developed PHVAC,k Rd,on γon ≤ Don (24)
from ACHP as shown in Eq. (15), while the pump power consump-
tion is calculated based on the maximum flow of the pump as
shown in Eq. (16). This study assumes a pump efficiency of 60%, a PHVAC,k Rd,o f f γo f f ≤ Do f f (25)
maximum pump flow rate of 1 kg/s, and a maximum delivery head
of 5 m based on products available on the market. PHVAC,k Rd,any γany ≤ Dany . (26)

Q˙ chil l er The variable J is the monthly utility cost for cooling ($), k is the
Pchil l er = (15) stage, N is the number of stages in the monthly billing period, Re
COPactual
is the energy charge rate ($/kWh), t is the time interval (h), D
 3
Ppump m˙ sc is the target demand cost for the monthly billing period ($), Rd is
= (16) the demand charge rate ($/kW), x0 is the initial state of charge of
Ppump,max m˙ max
the storage tank (-), xN is the storage state of charge at the end of
Ppump,max = η pump m˙ max ghmax (17) the simulation (-), and γ is a binary switch for demand charge (-)
with a value of 1 when the demand charge is applied for a time
PHVAC = (Pchil l er + Ppump ) (18) step and a value of 0 when the demand charge is not applied for
a time step. Note that the power consumption at stage k, PHVAC,k ,
The variable PHVAC is the HVAC power consumption (kW), Pchiller is itself a function of the storage charging/discharging rate at that
is the power consumption of the chiller (kW), Ppump is the power stage, Q˙ tank,k .
consumption of the pump (kW), Ppump,max is the maximum power The subscripts “on”, “off”, and “any” in this formulation refer
consumption of the pump (kW), m˙ max is the maximum mass flow to on-peak, off-peak, and anytime demand charges. The variable D
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 105

Fig. 7. Ice storage system solution flow diagram with inner loop solution using energy balance equations.

is used to place limits on the hourly power consumption at each

stage in the optimization as shown in Eqs. (24)–(26). For utility
rates without any demand charges, there are no limits on the de-
mand cost and so the demand charge rate Rd is set to zero for all
3 cases. For utility rates with only an on-peak demand, the terms
with subscripts “off” and “any” are zero (i.e. there is no limit), and
so on for the cases of an anytime demand or separate on and off-
peak demand charges. The constraints on the charging and dis-
charging rates of storage in Eq. (20) are determined by the mass
flow rate as shown in Eq. (2). These limits are also a function of
the charging and discharging heat transfer effectiveness of the stor-
age, which are determined through the state of charge and mode
of operation shown in Fig. 5. The cooling provided by the chiller
is constrained between zero and the chiller capacity as shown in
Eq. (21). The lower and upper limits on storage state of charge
in Eq. (22), xmin and xmax , are assumed to be 0 and 1 such that
only the latent capacity of the storage tank is considered. Finally,
Eq. (23) ensures a steady periodic solution.
The power consumption used in this monthly optimization
problem is the HVAC power and doesn’t include non-cooling elec-
Fig. 8. Schematic describing the optimal control problem.
trical usage. The addition of non-cooling electrical usage (lights,
appliances, etc.) in the cost function does not change the optimal
charging and discharging control of storage if this electrical usage
is constant over time. It also has minimal effect if the non-cooling the constraints are violated. For utility rates with demand charges,
electrical use is small compared to the HVAC power at times when the optimal utility cost J∗ is determined by solving the N variable
the target demand costs are established. This is a reasonable as- optimization problem for each iteration of an outer loop optimiza-
sumption for residential buildings that are unoccupied during the tion problem to determine the optimal values of Don, Doff, and Dany
weekday daytime periods when on-peak periods typically occur. as shown in Fig. 8.
The decision variables of this monthly optimization problem are This study used the fmincon function in MATLAB to solve the
the charging and discharging rates of the storage for all stages outer loop optimization problem. In order to illustrate the behav-
of the optimization period and the three monthly target demand ior of the optimal controller, daily simulation results are presented
costs. The optimization is performed over each month beginning under a TOU only residential utility rate structure offered in Miami,
at 0:00 on the first day and ending at 24:00 on the last day of the Florida (see Table 1) and using weather data for the month of May
month (day 30). Perfect load and weather forecast over the sim- in Miami. An ice storage system consisting of a 3-ton chiller and
ulation period are assumed. The problem is solved using a hybrid 125 gallons of ice storage was considered. Two consecutive days
method of dynamic programming [3] and non-linear optimization. are shown in the results to demonstrate the optimal controller’s
Dynamic programming utilizes Bellman’s principle of optimality performance for both charging and discharging the storage. Fig. 9
and guarantees a global minimum. At any time stage k, there are shows the chiller cooling rate as well as the building cooling load,
multiple paths to reach the previous time stage k-1. Each stage rep- and Fig. 10 shows the storage state of charge.
resents an hour time step in the simulation period, and each state It can be seen in Fig. 9 that the optimal controller minimizes
represents a state of charge in the storage tank. The state vari- chiller operation during the on-peak period on both days. The op-
ables are discretized from 0 to 1 in intervals of 0.01 and estab- erating cost for the month can be reduced by shifting the load to
lish a grid of possible paths between consecutive stages. Penalty off-peak periods with lower energy charges. The optimal controller
functions are utilized to handle constraints that would not allow also charges the storage by operating the chiller at a high capac-
transitions between states within the grid. Dynamic programming ity and meets the cooling load directly with the chiller when it
calculates the cost of all possible paths to find the least cost so- is not charging the storage. The storage state of charge shown in
lution for each consecutive stage in time. This process is used re- Fig. 10 reflects the chiller charging and discharging rate, where the
cursively to determine the minimum cost for the entire simula- storage is charged during the off-peak period and discharged dur-
tion period. Eq. (23) is satisfied by considering a specified state of ing the on-peak period. The optimal controller charged the storage
charge at the beginning and end of the time period over which the to 80% on the first day, but charged the storage fully (to a state of
optimization is conducted. The remaining constraints are enforced charge of 100%) before the hottest day of the month because of the
using penalty functions, which assign a very large cost to a path if higher cooling load forecasted for that on-peak period.
106 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

Table 1
Sample residential utility rate.

TOU energy only

On-peak energy Mid peak energy Off-peak energy

($/kWh) ($/kWh) ($/kWh)

Florida Power & Light 0.184 (<10 0 0 kWh) n/a 0.035(<10 0 0 kWh)
(FL) 0.204 (>10 0 0 kWh) 0.055(>10 0 0 kWh)
(12–8pm)

tioned in the description of the system model (see Section 3), we

set the maximum allowable state of charge of the storage tank to
be 0.9 and minimum allowable state of charge to be 0.1 to avoid
decreases in heat transfer effectiveness.
The control logic for charging the storage tank is shown in
Fig. 11. The storage tank will be charged during off-peak hours for
all utility rates. If there is no off-peak demand charge, the chiller
charges the storage using a full-capacity charging strategy. This
means the chiller operates at maximum capacity until the storage
reaches the desired charged level and then meets any loads with
only the chiller. If there is a demand charge during the off-peak
period, the controller utilizes a load forecast to calculate a mini-
mum rate to charge the storage, meaning the chiller charges the
storage at a constant rate that is just sufficient to fully charge the
storage at the end of the off-peak period. This approach tends to
minimize the peak power demand for charging. The chiller charg-
Fig. 9. Example of chiller operation using the optimal controller over two consecu- ing rate using the load-limiting charging strategy is calculated by
tive days under a TOU energy only utility rate in Miami, Florida. On-peak period is
to f f
represented by the gray shaded region.
i=1
Q˙ o f f,i t + Qcharge
Q˙ chil l er,o f f = , (27)
to f f

where Q˙ o f f is the building cooling load during the off-peak period

(kW), Qcharge is the amount of energy required to charge the stor-
age to the desired state of charge, and toff is the duration of the
off-peak period (h). Load forecasts for the off-peak period can be
updated over time during the charging period along with the en-
ergy required to charge the storage.
The charging rate, Q˙ chil l er,o f f , can achieve a minimum demand
during each off-peak period but is unnecessary for days when the
power consumption does not reach the peak demand. Therefore, a
target demand cost, Dk , is used to minimize the off-peak or any-
time demand cost of the month. This is reset to zero at the start
of each billing period and reset throughout the billing period as
the maximum of the current chiller load and previous target limit.
The update of the target demand cost at each time step can be
described as,
Fig. 10. Example of storage state of charge using the optimal controller over two
consecutive days under a TOU energy only utility rate in Miami, Florida. On-peak
Dk = Max(PHVAC,k Rd,any γany , PHVAC,k Rd,o f f γo f f , PHVAC,k Rd,on γon , Dk−1 ),
period is represented by the gray shaded region. (28)
where Dk is the target demand cost that is updated at each hourly
5. Generalized rule-based controller time step. By considering the overall demand cost ($) instead of
the power demand ($/kW), the controller can appropriately weigh
In this section, we present a generalized rule-based controller the difference between on-peak and off-peak demand charges. The
for residential buildings. The proposed strategy switches between chiller cooling rate is determined by the maximum of the demand-
different control strategies for charging and discharging the ice limit and the load-limiting charging strategy. This allows the con-
storage tank based on a given utility rate structure and is loosely troller to charge the storage faster on days when the peak demand
based on the method presented in the ASHRAE Handbook of HVAC cost target is not reached. The controller can then meet the re-
Applications [18]. The logic used by the generalized rule-based maining loads during the off-peak period at a higher chilled water
controller is inspired by the performance of the optimal controller supply temperature, which increases the system’s efficiency.
and will maximize the use of storage during the on-peak period. The control logic for discharging storage is depicted in Fig. 12.
The proposed control logic is designed to first determine the util- The controller uses load-limiting storage priority during the
ity rate period: on-peak, off-peak, or mid-peak. For utility rates on-peak period to maximize the use of storage. If the utility rate
that have a flat energy and anytime demand rate, an “effective” includes a mid-peak period, the controller prioritizes the use of
on-peak period is determined as the hours during which the cool- storage during the on-peak period, and the remaining storage is
ing loads exceed 0.75 of the chiller’s maximum capacity. As men- discharged during the mid-peak period using load-limiting storage-
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 107

Fig. 11. Control logic for charging storage using the generalized rule-based controller.

Fig. 12. Control logic for discharging storage in the generalized rule-based controller.

priority. The on-peak chiller loading is determined using Eq. (1), This is because the chiller has a lower efficiency when charging
and the mid-peak chiller loading is determined using Eq. (29). the storage, and so reducing the amount of storage discharged can
tmid ton reduce the overall energy cost.
i=1
Q˙ mid,i t + j=1
Q˙ on, j t − Qstorage
Q˙ chil l er,mid = (29)
tmid 6. Controller evaluation

The variable Q˙ chil l er,mid is the rate of cooling from the chiller In this section, the monthly performance of the generalized
during the mid-peak period (kW), Q˙ mid is the cooling load during rule-based controller is evaluated through comparisons with the
the mid-peak period (kW), Q˙ on is the cooling load during the on- optimal benchmark controller and the simpler heuristic controllers
peak period (kW), Qstorage is the available storage capacity (kWh), previously developed for commercial buildings. Daily comparisons
and tmid is the duration of the mid-peak period (h). between the generalized rule-based controller and the optimal
If there is no demand charge during the off-peak period, the controller are then used to better understand the differences be-
controller uses chiller-priority to preserve the storage for the tween the two controllers. An ice storage system consisting of a
on-peak period. If there is an off-peak demand charge applied, 3-ton chiller and 125 gallons of ice storage was used in the sim-
then the controller chooses between chiller-priority and demand- ulations. The simulations assumed perfect weather forecasts and
limiting control. Demand-limiting control limits the system power used weather data from TMY3 for ambient conditions and build-
consumption based on the target demand cost described in the ing loads [20]. The simulations used hour-long time steps for all
charging logic, in order to avoid incurring a higher anytime or off- control strategies. In this paper, we consider only the electricity
peak demand cost that month. The controller operates the chiller consumption due to space cooling; other non-cooling electricity
at a constant rate determined by the target demand cost, and consumption, such as lighting and appliance use, can vary signifi-
the storage is discharged if the chiller cannot meet the cooling cantly across households and was not included. However, the non-
load. The controller utilizes the cooling load forecast to evaluate cooling electrical loads are generally small during a typical day-
whether demand-limiting control will be able to meet the entire time weekday on-peak period when the residence is unoccupied
cooling load before the start of the on-peak period. If the con- and typically small in relation to the peak power associated with
troller will not be able to meet the off-peak cooling load with the cooling system. Thus, it is expected that the optimal charging
demand-limiting control, then chiller-priority will be used to en- and discharging control of the ice storage would not change sig-
sure all cooling load is met. If the utility rate has a flat energy nificantly if non-cooling electrical usage were included. In order to
rate and no demand charge, the controller utilizes chiller-priority account for the effect of residential occupancy on the operation of
because there is no economic benefit of discharging the storage. the proposed cooling system, the indoor set-point temperature was
108 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

Table 2
Utility rate parameters considered in monthly controller evaluation.

Parameter Range of values

Off-peak energy charge ($/kWh) 0.06

On-peak to off-peak energy charge 1 (Flat energy)
ratio 3 (TOU energy)
5 (TOU energy)
Demand charge ($/kW) R d,on R d,off Rd,any
20 10 0
20 0 0
0 0 10
0 0 0
On-peak duration (Hours) 4, 8, 12

defined as 76°F from 10 am to 4 pm on weekdays and 72°F the rest

of the time.

6.1. Monthly performance evaluation of the generalized rule-based

controller

The monthly performance of the generalized rule-based con-

troller was evaluated using a range of utility rates for two geo-
graphic locations with different climates: Miami, Florida and In-
dianapolis, Indiana. While these two climates have very similar
peak cooling loads, the monthly cooling load is much lower in
Indianapolis. The monthly performance, as well as that of the
other heuristic controls strategies discussed earlier, were compared
against the optimal controller. Weather data from TMY3 for the
month of May was used for Miami, Florida, and the month of June
was used for Indianapolis, Indiana. These months provide repre-
sentative cooling requirements for the cooling season in each lo-
cation. The performances of the different control strategies were
evaluated over month-long simulations using a range of different Fig. 13. Comparison of monthly operating costs (normalized by monthly cooling
utility rates created by combining different energy charges, de- load) between the generalized rule-based and optimal controller over all combina-
mand charges, and on-peak period durations. All combinations, 64 tions of evaluation parameters.
in total, of the parameter ranges shown in Table 2 were employed.
These cover the range of currently available utility rates for resi-
dential buildings [14]. edge about future weather and loads. While both controllers had
Values of 1, 3, and 5 were considered for the ratio of the on- perfect forecasts in the simulations, the duration of the forecasts
peak to off-peak energy charge, where a ratio of 1 represents a flat was different. For these simulations, the generalized rule-based
energy rate since the energy charge is constant. The off-peak en- controller utilized a perfect 24-hour forecast whereas the optimal
ergy charge was chosen to be $0.06/kWh based on available utility controller used a perfect forecast for the entire month. It is impor-
rates. The demand charge parameter includes no demand, anytime tant to note, however, that the performance of the optimal con-
demand, on-peak demand, and on-peak and off-peak demand. Two troller will be more sensitive to uncertainties in forecasts than the
different values of on-peak and off-peak demand were considered generalized rule-based controller since it requires monthly fore-
in this evaluation, resulting in 4 different combinations of demand casts whereas the generalized rule-based controller only utilizes
charge. The first case had an on-peak demand charge of $20/kW forecasts that are at most the length of the off-peak or on-peak pe-
and an off-peak demand charge of $10/kW, the second case had riod (typically less than 12 h). While it is unrealistic to expect ac-
an on-peak demand charge of $20/kW with no off-peak demand curate forecasts over a month-long period, the accuracy of 12-hour
charge, the third case had an anytime demand of $10/kW, where forecasts would be much better. Also, forecasts during off-peak and
the on-peak and off-peak charges were combined into 1 single de- on-peak periods can be updated over time within an actual imple-
mand charge, and the fourth case had no demand charge at all. mentation, which would improve overall performance. Therefore,
Values of 4, 8, and 12 h were considered for the duration of the performance differences between the optimal and heuristic con-
on-peak period. trollers would be smaller than those presented here.
The normalized monthly operating costs of the generalized Another key difference between the proposed generalized rule-
rule-based controller are compared against the optimal controller based control strategy and the optimal controller is in their ob-
in Fig. 13. The results were normalized by the monthly cooling load jectives. The generalized controller is designed to minimize the
of each location and are presented as the operating cost per unit of monthly demand charge while minimizing individual daily en-
cooling ($/ton-h). Results are presented in this normalized manner ergy costs subject to demand constraints, whereas the optimal
so that they can be compared across different climates and readily controller minimizes the overall operating cost by evaluating the
scaled for larger implementations. The results show that the gen- trade-off between energy and demand costs. For example, the gen-
eralized rule-based controller was always able to achieve operat- eralized rule-based controller limits the storage state of charge
ing costs within 20% of the optimal controller, and in many cases, based on its heat transfer effectiveness, but allowing the storage
within 10%. The discrepancy between the generalized controller to be fully charged can further reduce the operating costs during
and the optimal controller is largely due to differences in knowl- days with high cooling load. The optimal controller is able to as-
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 109

Table 3
Simulation parameters for cases in which the generalized rule-based controller exceeded 15% of
the minimal cost.

Parameter Simulation cases

Location FL FL FL FL IN IN

On-peak to off-peak energy charge ratio 1 5 5 5 1 5

On-peak demand charge ($/kW) 20 0 0 0 20 0
Off-peak demand charge ($/kW) 0 0 0 0 10 0
Anytime demand charge ($/kW) 0 10 10 0 0 10
Duration 8 8 12 12 12 12
Relative cost 1.155 1.157 1.168 1.151 1.164 1.171

Fig. 15. Comparison of chiller operation under generalized rule-based and optimal
control over two consecutive days using a TOU energy only utility rate in Miami,
Florida. ($0.06/kWh off-peak and $0.30/kWh on-peak, 4 h on-peak duration).

Fig. 14. Frequency distribution of monthly operating costs (normalized by monthly

cooling load) relative to the optimal controller for the generalized rule-based con- Table 4
troller, load-limiting storage-priority, and chiller priority over all combinations of Monthly cost comparison between controllers under TOU energy rate only in Mi-
evaluation parameters. ami, Florida.

Optimal controller Generalized rule-based Relative

($/ton-h) controller ($/ton-h) cost

Energy cost 0.0686 0.0708 1.032

sess this trade-off because it can identify the hottest days of the
month using a month-long forecast.
Fig. 14 shows the frequency distribution of the normalized
monthly operating costs of the various heuristic control strate- 6.2. Daily comparison between the generalized rule-based and
gies relative to the optimal controller. The generalized rule-based optimal controller
controller outperformed the chiller-priority and load-limiting
storage-priority control strategies. Load-limiting storage-priority Next, daily comparisons of the two controllers are used to high-
and chiller-priority have relative operating costs exceeding 50% of light the differences in system operation that result from the two
the optimal controller in many cases. These existing control strate- approaches. Two sample cases with different utility rate structures
gies are penalized for their full capacity charging strategy under and performance are presented. The first comparison is a case in
utility rates with anytime or off-peak demand. On the other hand, which the generalized rule-based controller performed within 5%
the generalized rule-based controller was able to achieve operating compared to the optimal controller. A utility rate with a $0.06/kWh
costs within 20% of the optimal controller, and the performance off-peak energy charge, $0.30/kWh on-peak energy charge, no de-
across all simulations was better than the existing heuristic con- mand charges, and an on-peak duration of 4 h was simulated in
trol strategies, which had operating costs 50% greater the optimal Miami, Florida. Fig. 15 presents the chiller cooling rates for the two
controller in some cases. controllers and an overlay of the cooling load profile, and Fig. 16
Table 3 shows the geographic locations and combinations of presents storage state of charge comparisons. The white bars repre-
utility rate parameters for cases in which the generalized rule- sent the generalized rule-based controller and the black bars rep-
based controller exceeded 15% of the minimal cost. These cases resent the optimal controller. Two consecutive days are shown to
have on-peak durations of either 8 or 12 h, and with the exception highlight the differences in controller behavior, and the second day
of one case, include demand charges. The combinations shown in shown is the hottest day of the month. The shaded areas in the fig-
Table 3 also include cases with TOU energy rates where the on- ures denote the on-peak period of each day. The operating cost for
peak energy charge is five times the off-peak charge. Most avail- this comparison is shown in Table 4.
able residential utility rates have shorter on-peak periods, and the As shown in Fig. 15, the two controllers follow a similar pattern
on-peak energy charge is typically lower for rates with long on- for chiller operation, which minimizes the use of the chiller dur-
peak periods. In other words, the generalized rule-based controller ing the on-peak period and charges the storage at a high capac-
performed best (within 15% of the optimal controller) for the most ity during the off-peak period. While the generalized rule-based
realistic utility rate structures. controller charges the storage at full capacity at the start of the
110 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111

Fig. 17. Comparison of chiller operation under generalized rule-based and optimal
Fig. 16. Comparison of ice storage state of charge under generalized rule-based and control over two consecutive days using a TOU energy with on-peak and off-peak
optimal control over two consecutive days using a TOU energy only utility rate in demand utility rate in Miami, Florida. ($0.06/kWh off-peak energy, $0.18/kWh on-
Miami, Florida. ($0.06/kWh off-peak and $0.30/kWh on-peak, 4 h on-peak dura- peak energy, $10/kW off-peak demand, $20/kW on-peak demand, 8 h on-peak du-
tion). ration).

off-peak period (hour 20), the optimal controller chooses to charge

the storage in the early morning (i.e. hours 24–31 in this example).
The optimal controller optimizes the charging scheme by charg-
ing the storage when the ambient temperature is lower and the
chiller has higher efficiency. When discharging storage, the gen-
eralized rule-based controller does not operate the chiller. This is
because it utilizes load-limiting storage priority, and since the on-
peak period only lasts for 4 h in this utility rate, the storage tank is
sufficiently charged so as to enable full storage control during the
on-peak period. The optimal controller actually operates the chiller
more than the generalized rule-based controller on both days as
shown in Fig. 15. This is unexpected because systems usually have
a lower operating cost if they are able to shift more load to the off-
peak period. However, the optimal controller is still able to achieve
a lower operating cost, and we understand this difference in op-
erating cost by considering the storage state of charge, shown in
Fig. 16.
Fig. 18. Comparison of ice storage state of charge under generalized rule-based
The optimal controller and generalized rule-based controller and optimal control over two consecutive days using a TOU energy with on-peak
both try to maximize the use of storage so that it is nearly de- and off-peak demand utility rate in Miami, Florida. ($0.06/kWh off-peak energy,
pleted by the end of each day. The generalized rule-based con- $0.18/kWh on-peak energy, $10/kW off-peak demand, $20/kW on-peak demand, 8 h
troller is limited to discharging to a 0.1 state of charge to avoid on-peak duration).
a decrease in heat transfer effectiveness, while the optimal con-
troller is allowed to discharge the storage completely. Furthermore,
while the generalized rule-based controller charges the storage to days with high cooling loads are small. This is also evidenced in
a 0.9 state of charge every off-peak period, the optimal controller the monthly operating cost of both controllers, where the general-
is allowed to decide the state of charge at the end of the off-peak ized rule-based controller is able to achieve a cost within 5% of the
period. This is shown in Fig. 16, where the optimal controller only optimal one.
charges the storage to a 0.4 state of charge on the first day and a The second comparison uses a utility rate structure that has a
0.8 state of charge on the hottest day. While this provides less stor- TOU energy rate with on-peak and off-peak demand simulated in
age capacity for the optimal controller during the on-peak period, Miami, Florida. This utility rate structure has an off-peak energy
the optimal controller is able to forecast over the whole month, charge of $0.06/kWh, an on-peak energy charge of $0.18/kWh, an
and knows the on-peak cooling load for this day is relatively low. on-peak demand charge of $20/kW, an off-peak demand charge of
This also means the optimal controller spends less time operating $10/kW, and an on-peak duration of 8 h. The results are shown in
the system in charging mode. As shown in the chiller performance Figs. 17 and 18 in a similar format as the previous example.
map, the chiller COP decreases as the chilled water supply tem- The two controllers have more differences in chiller operation
perature decreases. Therefore, the optimal controller is operating under this utility rate structure. When discharging the storage on
the chiller at a higher efficiency more frequently than the gener- the hottest day (i.e. hours 24–48), the optimal controller operates
alized rule-based controller. The optimal controller is able to evalu- the chiller at a low capacity during the second hour of the on-peak
ate the trade-off between the additional storage capacity for the on- period instead of operating the chiller at a constant capacity (see
peak period and the efficiency penalty when charging storage. Despite Fig. 17). While both controllers minimize the use of the chiller dur-
this difference, the storage state of charge is very similar for both ing the on-peak period, the optimal controller also minimizes the
controllers on the hottest day of the month. So while the optimal use of the chiller during the hottest hour of this on-peak period.
controller can consider additional trade-offs compared to the gen- The optimal controller is able to achieve a lower on-peak demand
eralized rule-based controller, the performance differences on the and energy cost. Table 5 shows that the optimal controller outper-
 A. Tam, D. Ziviani and J.E. Braun et al. / Energy & Buildings 197 (2019) 99–111 111

Table 5
Monthly cost comparison between controllers under TOU energy rate with on-peak and off-peak de-
mand in Miami, Florida.

Optimal controller Generalized rule-based Relative

($/ton-h) controller ($/ton-h) cost

Energy cost 0.1004 0.1081 1.08

On-peak demand cost 0.0203 0.0271 1.34
Off-peak demand cost 0.0343 0.0361 1.05
Total cost 0.1550 0.1713 1.11

forms the generalized rule-based controller over the entire month, Acknowledgement
and the biggest difference is in the on-peak demand cost.
Fig. 18 shows that both controllers have similar states of charge The authors would like to thank the Center for High Perfor-
on the first day, but the optimal controller charges the storage mance Buildings at Purdue University for their support of the re-
completely before the hottest day. The generalized rule-based con- search presented here.
troller is designed to limit the state of charge to 0.9 to avoid the
penalty of a decrease in heat transfer effectiveness discussed ear- References
lier. The optimal controller is able to achieve a lower on-peak de-
[1] Bell, I.H., 2012. ACHP documentation. 1.4 [WWW Document]. URL http://achp.
mand cost for the month because it is able to optimally evaluate sourceforge.net/index.html.
the trade-off between the heat transfer effectiveness penalty and ad- [2] I.H. Bell, J. Wronski, S. Quoilin, V. Lemort, Pure and pseudo-pure fluid ther-
ditional storage capacity for the on-peak period. mophysical property evaluation and the open-source thermophysical property
library coolprop, Ind. Eng. Chem. Res. 53 (2014) 2498–2508. https://doi.org/10.
1021/ie4033999.
[3] R. Bellman, Dynamic Programming, Princeton University Press, Princeton, N.J,
1957.
7. Conclusion [4] J.E. Braun, A near-optimal control strategy for cool storage systems with dy-
namic electric rates, HVAC&R Res. 13 (4) (2007) 557–580.
In this study, a generalized rule-based control strategy was de- [5] W.J. Cole, T.F. Edgar, A. Novoselac, Use of model predictive control to enhance
the flexibility of thermal energy storage cooling systems, in: Presented at the
signed for residential cooling systems with integrated TES for util-
2012 American Control Conference (ACC), 2012, pp. 2788–2793. https://doi.org/
ity rates structures available today to residential customers. A sys- 10.1109/ACC.2012.6314689.
tem model consisting of individual models for an ice storage tank, [6] K. Drees, J. Braun, Development and evaluation of a rule-based control strategy
a packaged chiller, secondary loop, and building load was created for ice storage systems, HVACR Res. 2 (1996) 312–334. https://doi.org/10.1080/
10789669.1996.10391352.
in order to conduct simulations for different sample residential [7] A. Hajiah, M. Krarti, Optimal control of building storage systems using both ice
variable utility rates. The rule-based control strategy was moti- storage and thermal mass – Part I: simulation environment, Energy Convers.
vated by the performance of an optimal controller, which used Manag. 64 (2012) 499–508. https://doi.org/10.1016/j.enconman.2012.02.016.
[8] A. Hajiah, M. Krarti, Optimal controls of building storage systems using both
dynamic programming to minimize the total energy and demand ice storage and thermal mass – Part II: parametric analysis, Energy Convers.
costs over monthly billing periods. The generalized rule-based con- Manag. 64 (2012) 509–515. https://doi.org/10.1016/j.enconman.2012.02.020.
troller was compared in simulations against an optimal controller [9] G.P. Henze, Parametric study of a simplified ice storage model operating under
conventional and optimal control strategies, J. Sol. Energy Eng. 125 (2003) 2.
as well as existing heuristic control strategies from the literature. https://doi.org/10.1115/1.1530629.
The control strategies were evaluated across a matrix of differ- [10] F. Kung, M. Deru, E. Bonnema, Evaluation Framework and Analyses for Thermal
ent residential utility rate structures for two geographic regions. In Energy Storage Integrated with Packaged Air Conditioning, National Renewable
Energy Laboratory (NREL), Golden, CO, 2013.
comparison to the optimal controller, the generalized rule-based [11] E.W. Lemmon, I.H. Bell, M.L. Huber, M.O. McLinden, NIST Standard Reference
controller performed much better than the other heuristic con- Database 23: reference Fluid Thermodynamic and Transport Properties-REF-
trol strategies. The generalized rule-based controller was able to PROP, Version 10.0, National Institute of Standards and Technology. Standard
Reference Data Program, 2018 2018.
achieve operating costs that were within 20% of the optimal cost,
[12] C.-C. Lo, S.-H. Tsai, B.-S. Lin, Ice storage air-conditioning system simulation
while load-limiting storage-priority and chiller-priority had a sig- with dynamic electricity pricing: a demand response study, Energies 9 (2016)
nificant number of cases with costs that were 50% greater than the 113. https://doi.org/10.3390/en9020113.
optimal cost. Furthermore, the generalized rule-based controller [13] N. Luo, T. Hong, H. Li, R. Jia, W. Weng, Data analytics and optimization of an
ice-based energy storage system for commercial buildings, Appl. Energy 204
achieved costs within 10% of the optimal cost in over 50% of the (2017) 459–475.
simulation cases, and had better performance under rates that are [14] NREL, 2018. U.S. Utility Rate Database [WWW Document]. URL https://openei.
more similar to currently available residential utility rates. Even org/apps/USURDB/
[15] F. Sehar, M. Pipattanasomporn, S. Rahman, Coordinated control of building
though the generalized rule-based controller led to higher oper- loads, PVs and ice storage to absorb PEV penetrations, Int. J. Electr. Power En-
ating costs than the optimal controller, it only requires weather ergy Syst. 95 (2018) 394–404. https://doi.org/10.1016/j.ijepes.2017.09.009.
forecasts over the discharge period and is relatively easy to imple- [16] Sun, C., Temple, K., Rossi, T., Braun, J.E., 2006. Interaction between dynamic
electric rates and thermal energy storage control, Final Report for RP-1252.
ment. The generalized rule-based controller could be improved by [17] Y. Sun, S. Wang, F. Xiao, D. Gao, Peak load shifting control using different
considering trade-offs that are evaluated by the optimal controller, cold thermal energy storage facilities in commercial buildings: a review, En-
such as energy costs versus demand costs, and heat transfer effec- ergy Convers. Manag. 71 (2013) 101–114. https://doi.org/10.1016/j.enconman.
2013.03.026.
tiveness penalty versus additional storage capacity for the on-peak [18] , in: ASHRAE Handbook of HVAC Applications, 2011, p. 42. 1-42.39.
period. However, these improvements would likely add complexi- [19] J. West, J. Braun, Modeling partial charging and discharging of area-constrained
ties that impact the overall economics of implementing the control ice storage tanks, HVACR Res. 5 (1999) 209–228. https://doi.org/10.1080/
10789669.1999.10391234.
strategy. Future work will include evaluating the sensitivity of the
[20] S. Wilcox, W. Marion, Users Manual for TMY3 Data Sets (Revised), National
generalized rule-based controller to forecast uncertainty. Renewable Energy Lab.(NREL), Golden, COUnited States, 2008.
[21] Z.(Jerry) Yu, G. Huang, F. Haghighat, H. Li, G. Zhang, Control strategies for in-
tegration of thermal energy storage into buildings: state-of-the-art review, En-
ergy Build. 106 (2015) 203–215. https://doi.org/10.1016/j.enbuild.2015.05.038.
Conflict of interest [22] Yudong Ma, F. Borrelli, B. Hencey, B. Coffey, S. Bengea, P. Haves, Model predic-
tive control for the operation of building cooling systems, IEEE Trans. Control
Syst. Technol. 20 (2012) 796–803. https://doi.org/10.1109/TCST.2011.2124461.
None.