Applied Energy 203 (2017) 959–971

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Optimized dispatch in a first-principles concentrating solar power

production model
Michael J. Wagner a, Alexandra M. Newman b,⇑, William T. Hamilton b, Robert J. Braun b
a
National Renewable Energy Laboratory, Thermal Systems Group, 15013 Denver West Parkway, Golden, CO 80401, United States
b
Colorado School of Mines, Department of Mechanical Engineering, 1500 Illinois Street, Golden, CO 80401, United States

h i g h l i g h t s

A profit maximizing solution operates concentrated solar power towers.

Dispatch optimization integrates a techno-economic analysis tool.

An optimized strategy improves plant profitability by 5–20%.

Further improvements derive from a reduction in the number of cycles by 50%.

Power cycle start-ups reduce from 370 to 250 per year without affecting energy output.

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

Article history: Concentrating solar power towers, which include a steam-Rankine cycle with molten salt thermal energy
Received 25 November 2016 storage, is an emerging technology whose maximum effectiveness relies on an optimal operational and
Received in revised form 18 June 2017 dispatch policy. Given parameters such as start-up and shut-down penalties, expected electricity price
Accepted 19 June 2017
profiles, solar availability, and system interoperability requirements, this paper seeks a profit-
Available online 15 July 2017
maximizing solution that determines start-up and shut-down times for the power cycle and solar recei-
ver, and the times at which to dispatch stored and instantaneous quantities of energy over a 48-h horizon
Keywords:
at hourly fidelity. The mixed-integer linear program (MIP) is subject to constraints including: (i) mini-
Dispatch optimization
Grid integration
mum and maximum rates of start-up and shut-down, (ii) energy balance, including energetic state of
Concentrating Solar Power (CSP) the system as a whole and its components, (iii) logical rules governing the operational modes of the
Thermal energy storage power cycle and solar receiver, and (iv) operational consistency between time periods.
Mixed-integer linear programming The novelty in this work lies in the successful integration of a dispatch optimization model into a
Systems analysis detailed techno-economic analysis tool, specifically, the National Renewable Energy Laboratory’s
System Advisor Model (SAM). The MIP produces an optimized operating strategy, historically determined
via a heuristic. Using several market electricity pricing profiles, we present comparative results for a sys-
tem with and without dispatch optimization, indicating that dispatch optimization can improve plant
profitability by 5–20% and thereby alter the economics of concentrating solar power technology. While
we examine a molten salt power tower system, this analysis is equally applicable to the more mature
concentrating solar parabolic trough system with thermal energy storage.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Background

The ability of renewable energy to be dispatched flexibly

Abbreviations: CSP, Concentrating Solar Power; HTF, Heat Transfer Fluid; MIP, enables significant market penetration compared to renewable
Mixed-Integer Linear Program; MSPT, Molten Salt Power Tower; PPA, Power energy systems that are highly variable (e.g., wind) and/or that lack
Purchase Agreement; SAM, System Advisor Model; TES, Thermal Energy Storage;
associated storage systems (e.g., photovoltaics without storage).
TOD, Time-of-Dispatch.
⇑ Corresponding author. We examine one type of solar technology, concentrating solar
E-mail addresses: michael.wagner@nrel.gov (M.J. Wagner), anewman@mines. power (CSP), that manifests itself as: parabolic trough, linear fres-
edu (A.M. Newman), whamilton@mines.edu (W.T. Hamilton), rbraun@mines.edu nel, dish stirling, and power tower. The latter, and the one
(R.J. Braun).

http://dx.doi.org/10.1016/j.apenergy.2017.06.072
0306-2619/Ó 2017 Elsevier Ltd. All rights reserved.
960 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

addressed in this paper, is thought to possess the most significant and generic systems. The software is free to download and use,
potential for improvements in efficiencies and reductions in cost and the tools developed in the current work are freely available
[1]. Concentrating solar power tower technology uses thousands [11]. Each technology can be paired with a financial model to eval-
of sun-tracking mirrors (heliostats) that focus on a central receiver uate the economic performance of a project within particular mar-
to heat molten salt to temperatures above 565°C (1050°F). The ket, incentive, and cost environments.
molten salt can then be pumped to a power cycle to generate elec-
tricity or efficiently stored for use when sunlight is not available 1.1. Related work
[2]. However, the economic viability and widespread implementa-
tion of CSP technologies are strongly tied to their ability to extend Optimization modeling has been applied to many types of
their diurnal operational characteristics across peak demand time energy systems, e.g., [9] who retrofit an existing building and
periods and during periods when solar energy is curtailed due to determine a corresponding dispatch strategy, and [12] who exam-
the sun setting or cloud cover [3]. Thermal energy storage (TES) ine multiple objectives in optimizing stand-alone hybrid energy
is an enabling technology which can amass the energy captured systems, also with the corresponding dispatch. Other authors
by the receiver as a reserve for dispatch at a later, more economical examine only dispatch, e.g., [13], who apply a simulation model
time. In fact, TES integration enables CSP to be a dispatchable to a hybrid photovoltaic and tri-generation power system to
renewable resource whose economics are enhanced by both decrease waste from excess heat, while [14] formulate an opti-
improved utilization of the power cycle and an ability to shift mization model (a mixed-integer linear program, like ours) that
power production to better coincide with peak demands and combines both dispatchable and intermittent power, the latter as
high-value-electricity time periods [4]. a result of a virtual plant, to maximize profits. Similarly, [15]
High-temperature molten salt TES has been successfully imple- develop an optimization model that dispatches wind, but, in con-
mented in CSP tower systems [5,6] and in parabolic trough sys- trast to the previous work, theirs focuses on minimizing active
tems, the latter in an indirect manner through use of an power losses in the system while constraining reactive power;
intermediate oil-to-molten salt heat exchanger. So-called direct the model is solved heuristically. Thorin et al. [16–18] operate in
TES systems such as the power tower technology use molten salt a market environment (as does [14]), the former for a unit commit-
both as the storage medium and as the heat transfer fluid in the ment problem, applying an exact approach (i.e., Lagrangian Relax-
receiver, thereby avoiding the intermediate heat exchanger and ation) to a mixed-integer program; Cho et al. [17] optimize a
improving system efficiency and dispatchability [7]. combined cooling, heating, and power system to optimize the
The maximum storage capacity of the TES system is determined tradeoffs between system cost, energy production and emissions,
during a plant design process that considers several factors includ- and test their model on a variety of geographic sites in the U.S.
ing the thermal power rating of the solar field and power cycle sub- with differing weather conditions; Fürsch et al. [18] examine the
systems, plant location, project economics, and the desired capacity expansion of a power network and the corresponding dispatch
factor, which is defined as the quotient of total annual electrical strategies in Europe; using an optimization model which combines
energy production and the electrical energy production should both investment and dispatch decisions, they conclude that even
the plant operate continuously at rated power output. Thermal optimal grid extensions, coupled with capital cost reductions for
energy storage sizing also depends on the operational scheme. renewable technologies, leads to significantly higher overall aver-
For example, a plant that intends to operate primarily during age electricity system costs over a time horizon of three to four
high-revenue morning or evening periods while reducing produc- decades. Parisio et al. [19] use model predictive control within an
tion during daylight hours requires more TES capacity than a plant optimization (mixed-integer programming) framework in which
with an identical capacity factor that generates power during all the goal is to minimize costs subject to microgrid system con-
daylight hours. CSP plants that target dispatch during high- straints such a capacities, minimum up- and down-times, and
revenue periods operate differently than those that minimize the start-up and shut-down requirements. They test instances of their
average cost of energy. The former relies more extensively on a model on an experimental microgrid in Greece. Zheng et al. [20]
carefully planned dispatch schedule that anticipates the timing provide a review of bio-inspired optimization of sustainable energy
and level of thermal power production in the solar field, energy systems. These works examine problems similar to ours in that dis-
consumption for receiver and plant start up, and the charge state patch policies are considered, some even using the mathematical
of TES over time. Formal optimization methods can determine framework in this paper. However, none of these examines concen-
the dispatch profile that maximizes electricity sales revenue over trating solar power in particular, with its own sets of objectives
a particular time horizon given a specific system configuration, and rules. We next discuss the research specific to power tower
expected solar resource, pricing or time-of-dispatch (TOD) profile, technology.
and operational constraints – a process referred to as dispatch Simulation is used to predict the total electrical energy produc-
optimization. tion from an existing or previously designed CSP plant over its life-
The intelligent dispatch of stored energy can greatly enhance time in order to evaluate the financial return on investment, the
the value of electricity by providing firm capacity and ancillary ser- cost of energy, the environmental (mitigation) impact, or some
vices, and by generating electricity during time periods in which other measure of interest. The standard method for CSP simulation
rates are especially high [8]. Dispatch optimization involves the requires calculation of plant behavior over a time horizon (typi-
manipulation of the timing and rate at which electricity is gener- cally, one year with one-hour time steps) [21], and it develops a
ated by the power cycle and captures both physical processes picture of long-term energy production by sequentially modeling
and time [9]. This paper presents a methodology, implementation, performance at relatively short time steps compared to the overall
and publicly available tool for simulating CSP power tower systems time window of interest (e.g., hourly calculations to establish life-
with optimized dispatch. The method expands on previous work time metrics). CSP systems are primarily constrained by immediate
by directly incorporating formal optimization techniques into the concerns, such as component or subsystem operational states, con-
SAM [10] simulation software, for which previous research has servation of mass and energy, and heat transfer, thermodynamic,
relied on heuristics or on optimizing dispatch using simulation or thermo-mechanical principles.
output a posteriori as optimization model input. SAM assesses The previous dispatch approach implemented in SAM uses a
CSP performance, simulating renewable technologies including simple heuristic that allows the user to specify requirements
CSP, wind, geothermal, photovoltaic, biomass, solar hot water, before thermal storage can be dispatched; this heuristic does not
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 961

consider the expected future thermal energy production, TES easily represented in an energy-balance MIP model. For example,
charge state, or price at which electricity can be sold, but instead the mechanical stress associated with frequent thermal cycling of
determines the operational state of the power cycle based on the power generation equipment may lead to an increase in the fre-
current TES charge state and the hour of the day. The heuristic quency of required maintenance [27]. A detailed model can capture
can improve plant production during high-value hours as exempli- these thermo-mechanical impacts when the plant control is influ-
fied by SAM or Guédez et al. [22], but can ultimately decrease the enced by optimized dispatch scheduling.
utilization of the solar field throughout the year because of TES
over-charge situations. 1.3. Operating considerations
By contrast, this work adopts a formal approach by formulating
the problem as a mixed-integer program (MIP) that leverages The plant dispatch schedule determines the timing and produc-
state-of-the-art modeling languages and solvers ([23,24]) to make tion level from the power cycle (turbine, generator, condenser, and
the solution of a mathematical problem containing thousands of associated equipment). During operation, the power cycle con-
variables and constraints tractable. sumes stored thermal energy from the TES system. Thermal energy
Madaeni et al. [25] present a simplified approach for determin- storage is charged using high-grade thermal energy that is gener-
ing an optimal dispatch profile while implementing MIP tech- ated by the solar field during daytime operation, and energy gener-
niques. The authors use SAM to generate an hourly thermal ation is affected by the optical and thermal efficiency of the solar
power production profile throughout the year that is considered field, by the intensity of the available solar resource, and by the
as fixed input to the MIP model originally outlined in [26]. This operational state of the solar field. Receiver and power cycle
approach factors in the simulated performance of the solar field, start-up sequences are not necessarily coordinated, so both sys-
but omits interactions between the solar field and thermal storage tems may operate independently with shared interest only in the
or the power cycle. The latter subsystems are modeled as part of a energy state of the TES system. In some cases, the receiver must
MIP that determines the TES state of charge and electricity produc- curtail energy generation to avoid over-charging thermal storage
tion from the cycle. This method improves tractability by reserving (thus wasting solar energy).
the detailed model to generate fixed input while utilizing a simpli- Before the power cycle or receiver can produce electricity or
fied energy balance model to characterize TES charge and power thermal energy, respectively, start-up requirements must be satis-
cycle generation. Furthermore, [25] employ a rolling time horizon fied, including both a minimum start-up period and a minimum
methodology in which they consider a 48-h time horizon, updated energy state requirement which are surrogates for temperature
every 24 h. Our work largely adopts this approach, but considerations. In the latter case, the plant equipment cools during
importantly, uses the optimized schedule to control operational shutdown periods and must overcome the system’s thermal inertia
decisions within SAM’s detailed simulation model, whereas the to begin generating steam that powers the turbine. Likewise, the
work of [25] uses the results from the MIP as the actual estimate receiver consumes energy as it heats up and must complete a
of plant production throughout the year. start-up procedure before producing useful thermal energy. Fur-
thermore, turbine and heat exchanger equipment manufacturers
1.2. Goals of the current work limit the maximum rate of temperature increase during start-up
to avoid thermal stress and mechanical failure risks. Both the
Dispatch optimization improves the profitability of existing or energy and duration start-up requirements must be met before
planned CSP facilities, but it is also of great interest to policymak- equipment can begin producing power. These requirements are
ers and researchers who seek to better understand the projected implemented as a constraint on the maximum energy delivered
performance of CSP systems under various deployment and grid for start-up during any given time period. Although the duration
operations scenarios. However, previous work (cf. [8,25,26]) con- of start-up must last for at least a minimum number of time steps,
siders dispatchability from the perspective of grid integration in longer start-up durations are allowed in practice based on energy
which CSP systems are designed at an energy-flow and system siz- availability, and the model must provide this flexibility.
ing level to assess suitability for meeting grid and market Two start-up scenarios are possible for the power cycle: (i) cold
demands. The contribution of our work is its evaluation of the rela- start-up, which occurs when the power cycle has shut down for any
tionship between optimal dispatch profiles and technology design. period of time and seeks to restart; and (ii) hot start-up, which
Accordingly, this work fills the gap between prescriptive grid-level occurs when the power cycle has been in standby mode and seeks
models on the one hand that indicate desired technology perfor- to restart. Cold start-up requires an additional energy contribution
mance subject to high-level operational requirements (e.g., plant and incurs more component wear and tear, whereas hot start-up
start-up, maximum energy generation) and descriptive perfor- can happen immediately (from the perspective of the hourly
mance simulations on the other hand whose primary concern is model).
to dynamically synthesize expected plant productivity and Standby is a mode of operation in which a small (but non-
financial return given specific component or subsystem thermo- trivial) amount of thermal energy is consumed during each time
mechanical performance expectations. SAM develops these esti- period to maintain the power cycle and/or receiver equipment in
mates using annual ‘‘macro-simulations” that consist of thousands a hot state, ready to quickly ramp up for electricity generation;
of sequential ‘‘micro-simulations” within a time series, and the however, no electricity is produced in standby mode. Conse-
plant behavior at any given time step may depend on the state of quently, maintaining the power cycle in standby mode is of value
the system in the previous time step(s). The SAM molten salt if multiple start-up cycles are anticipated over a relatively short
power tower (MSPT) model is configured as illustrated in Fig. 1. time span, or if the energy penalty or ramp rate requirement for
The MIP in SAM operates under the following assumptions: (i) start-up is sufficiently severe to justify the small rate of energy
solar field thermal production over time is calculated using a sim- consumption by the power cycle.
plified ‘‘forecast” model and provided as a fixed input to the dis- The receiver can also operate in a standby mode during cloudy
patch model, and (ii) power cycle efficiency depends linearly on periods to avoid the full start-up procedure. In standby, salt from
thermal input to the cycle and on the ambient temperature, and the cold storage tank is pumped through the receiver, and the flow
these efficiency corrections can be implemented as independent is diverted back into the cold tank where the fluid temperature can
terms (see Section 3). Dispatch optimization enables investigation decay at a rate that corresponds to the thermal losses from the
of detailed plant performance issues that are too complex to be receiver. Finally, the model accounts for receiver shutdown energy
962 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

Fig. 1. Molten Salt Power Tower system configuration that is modeled in SAM. The system consists of a heliostat field, molten salt receiver, direct TES system, steam
generation system, Rankine power cycle, and heat rejection system (Graphic Ó NREL/Al Hicks).

consumption in which the heliostat field provides sufficient energy associated with cycle start-up, receiver start-up, and change in
to allow the salt to drain out without freezing before the solar field electricity production between time steps are subtracted from
ends operation for the day. The draining procedure requires the revenue.
approximately fifteen minutes while sunlight is still available,
ðRÞ maximize
and this effect is modeled as the consumption of 25% of the hourly Xh 
_ t  Lr ðxrt þ xrsu
D  Pt ct ð1  gct Þw rl rsb c _h r
energy used at the minimum receiver production rate. t þ Q y t Þ  L xt  W y t
t2T

2. Mathematical formulation _ b ycsb  ðW
W _ rsb þ Ehs =DÞyrsu  Ehs =D  yrsb  Ehs =D  yrsd
t t t t

ct ðC rsu yrsup

t þ C rhs yrhsp
t Þ
The parameters, sets, variables, objective function, and con- i
straining relationships are described in this section. The model ct ðC csu ycsup
t þ C chs ychsp
t þ aycsd
t þC
_ t Þ þ ct ðPDxrt þ ayrt Þ
dW d
w
takes the parameters and sets as given and determines values for
the decision variables to maximize an objective function while ð1Þ
adhering to the constraints. Some parameters and all variables
are subscripted with time t, indicating the time-varying nature of 2.4. Constraints
the decisions.
The relationships among the variables and parameters are
2.1. Parameters and sets established with a set of simultaneous equations and inequalities.
These constraints are presented below topically with a brief
The following MIP, (R), requires the initial operational state of description.
the system, the collector field and receiver energy generation pro-
file, the expected cycle conversion efficiency profile as a function of 2.4.1. Receiver operations
ambient temperature and thermal input, and the energy price or Receiver operations constraints include:
tariff profile (Table 1). (Initialization parameters used to set vari- Receiver Start-up
able values at t ¼ 0 follow variable notation and are not included
here.) ursu
t t1 þ D  xt
6 ursu rsu
8t 2 T : t P 2 ð2aÞ
ursu
t 6 Er yrsu
t 8t 2 T ð2bÞ
rsu
2.2. Variables u
yrt 6 t r þ yrt1 8t 2 T : t P 2 ð2cÞ
E
t þ yt1 6 1 8t 2 T : t P 2
yrsu ð2dÞ
r
The variables (see Table 2) describe energy (thermal kWt -hr or
electric kWe -hr) states and power flows (thermal kWt or electric xrsu 6 Q ru yrsu 8t 2 T ð2eÞ
t t
kWe ) in the system. Note that there is a one-to-one conversion
_ ‘‘u,” if Q in
t ¼ 0 then:
between kWt and kWe units. Continuous variables ‘‘x,” ‘‘w,”
and ‘‘s” representing power and energy relate to the receiver, yrsu
t ¼ 0 8t 2 T ð2fÞ
power cycle, and TES. Binary variables ‘‘y” enforce operational Receiver Supply and Demand
modes and sequencing such that start-up must occur before nor-
mal operation, for example. xrt þ xrsu rsd rsd
yt 6 Q in 8t 2 T
t þ Q t ð3aÞ

t yt 8t 2 T
xrt 6 Q in ð3bÞ
r
2.3. Objective function
xrt P Q rl yrt 8t 2 T ð3cÞ
The objective maximizes electricity sales, which are repre- if Q in ¼ 0 then:
t
sented as the summation over time of the product of electricity
yrt ¼ 0 8t 2 T ð3dÞ
price and power generation less parasitic losses. Cost penalties
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 963

Table 1
Parameters and sets used in (R).

Symbol Units Description

Sets
T Set of all time steps in the time horizon, T ¼ jT j
Time-indexed parameters
⁄
Q in kWt Power generated by the solar field in time t
t
Pt $/kWe-hr Electricity sales price in time t
W_ net kWe Net power production upper limit in time t
t
_ min
W kWe Minimum net power production in time t
t
gct – Normalized condenser parasitic loss in time t
ct – Exponential time weighting factor; CðtÞ , where C  0:99
⁄
Drs
t
– Estimated fraction of time step t used for receiver start-up
gamb
t
– Cycle efficiency adjustment factor in time t

Steady-state parameters
P
P – Mean sales price ($/kWe-hr); t2T P t =T
g des – Cycle nominal efficiency
⁄
gp – Slope of linear approximation of power cycle performance curve
s hr Frequency of optimization problem execution
Eu kWt-hr Energy storage capacity
Er kWt-hr Required energy consumed to start receiver
Ec kWt-hr Required energy consumed to start cycle
Ehs kWt-hr Heliostat field startup or shutdown parasitic loss
Wu kWe Cycle electric power rated capacity
Wl kWe Minimum electric power output from cycle
_h
W kWe Heliostat field tracking parasitic loss
_b
W kWe Power cycle standby operation parasitic load
_ rsb
W kWe Tower piping heat trace parasitic loss
Qu kWt Cycle thermal power capacity
Ql kWt Minimum operational thermal power input to cycle
Q ru kWt Allowable power per period for receiver start-up
Q rl kWt Minimum operational thermal power delivered by receiver
Q rsd kWt Required thermal power for receiver shut-down
Q rsb kWt Required thermal power for receiver standby
Qc kWt Allowable power per period for cycle start-up
Qb kWt Standby thermal power consumption per period
Lr kWe/kWt Receiver pumping power per unit power produced
Lc kWe/kWt Cycle Heat Transfer Fluid (HTF) pumping power per unit energy consumed
C rsu $ Penalty for receiver start-up (from 0)
C rhs $ Penalty for receiver start-up (from hot standby)
C csu $ Penalty for cycle start-up (from 0)
C chs $ Penalty for cycle start-up (from hot idle)
C dW $/kWe Penalty for any positive change in electricity production
D hr Time step duration
Dl hr Minimum duration of receiver start-up in period
a $ Conversion factor between unit-less and monetary values
M A sufficiently large number
⁄
Parameter is calculated from fixed input and discussed below.

Logic Governing Receiver Modes straint (2d) ensures that receiver start-up mode does not persist
while the receiver is operating in power-producing mode by disal-
t þ y t 6 1 8t 2 T
yrsu ð4aÞ
rsb lowing start-up in the time step following normal power produc-
tion operation. Constraint (2e) ensures that the actual power
yrt þ yrsb
t 6 1 8t 2 T ð4bÞ
used for receiver start-up is no more than the ramp rate limit for
t 6 yt1 þ yt1 8t 2 T : t P 2
yrsb ð4cÞ
r rsb
each time step. Constraint (2f) prevents receiver start-up from
yrsup
t P yrsu
t  yrsu
t1 8t 2 T : t P 2 ð4dÞ occurring in time periods with trivial solar resource.
The total power produced by the receiver has an upper bound of
yrhsp
t t1 Þ 8t 2 T : t P 2
P yrt  ð1  yrsb ð4eÞ
the available energy Q in
t , and any start-up or shutdown energy con-
yrsd
t1 P ðyrt1  yrt Þ þ ðyrsb
t1  t Þ
yrsb 8t 2 T : t P 2 ð4fÞ
sumption detracts from production according to Constraint (3a).
(R) considers receiver start-up inventory and the criteria that The receiver can only generate thermal power when it is in
must be satisfied in order for it to produce useful power. Constraint power-producing mode (i.e., yrt ¼ 1) by Constraint (3b). Constraint
(2a) tracks start-up energy ‘‘inventory” using an inequality, rather (3c) is enforced because of molten-salt pump operating limits and
than an equality, to allow inventory to reset to zero in time periods heat transfer requirements in the receiver, ensuring that the recei-
following start-up completion; inventory is naturally maximized ver energy generation must satisfy a minimum threshold. Con-
by the problem and can only be nonzero for time steps in which straint (3d) ensures that the receiver power-producing mode
the receiver is starting up by Constraint (2b). Constraint (2c) allows does not persist when no energy is available.
receiver power production only after start-up has been completed While the receiver is in standby mode, molten salt is circulated
or when the receiver was operating in the previous time step. Con- between the cold TES tank and receiver, enabling fast restart. A
964 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

Table 2
Variables used in (R).

Symbol Units Description

Continuous
xt kWt Cycle thermal power consumption at t
w_t kWe Electrical power generation at t
w_ dt kWe Positive change in electricity production at t
xrt kWt Thermal power delivered by the receiver at t
xrsu
t kWt Receiver start-up power consumption at t
ursu
t kWt-hr Receiver start-up energy inventory at t
ucsu
t kWt-hr Cycle start-up energy inventory at t
st kWt-hr TES reserve quantity at t (auxiliary variable)
Binary
yrt 1 if receiver is generating ‘‘usable” thermal power at time t; 0 otherwise
yrsu
t 1 if receiver is starting up at time t; 0 otherwise
yrsb
t
1 if receiver is in standby mode at time t; 0 otherwise
yrsd
t
1 if receiver shut down at time t; 0 otherwise
ycsu
t 1 if cycle is starting up at time t; 0 otherwise
ycsb
t
1 if cycle is in standby mode at time t; 0 otherwise
ycsd
t
1 if cycle is shutting down at time t; 0 otherwise
yt 1 if cycle is generating electric power at time t; 0 otherwise
yrsup
t
1 if receiver is starting up at time t and was not in standby mode at time t  1; 0 otherwise
yrhsp 1 if receiver is starting up at time t and was in standby mode at time t  1; 0 otherwise
t
ycsup
t
1 if cycle is starting up at time t and was not in standby mode at time t  1; 0 otherwise
ychsp 1 if cycle is starting up at time t and was in standby mode at time t  1; 0 otherwise
t

smaller hot start-up penalty is enforced when beginning normal Logic Governing Cycle Modes
operation from standby mode. Neither standby and start-up modes
(Constraint (4a)) nor standby and power-producing modes (Con- t þ yt1 6 1 8t 2 T : t P 2
ycsu ð7aÞ
straint (4b)) can coincide. Standby mode can persist over time, but ycsb
t t1 8t 2 T : t P 2
6 yt1 þ ycsb ð7bÞ
must follow time steps in which the receiver was either in standby
or power-producing mode (Constraint (4c)). Constraints (4d) and t þ yt
ycsu csb
6 1 8t 2 T ð7cÞ
(4e) enforce logic associated with incurring a penalty for receiver yt þ ycsb
t 6 1 8t 2 T ð7dÞ
ycsup t  yt1 8t 2 T : t P 2
start-up from an off or standby state, respectively. Constraint (4f) t P ycsu csu
ð7eÞ
enforces the logic for shut-down from a power producing or standby
state. Constraint (10a) ensures non-negativity for receiver start-up
ychsp
t t1 Þ 8t 2 T : t P 2
P yt  ð1  ycsb ð7fÞ
power consumption and receiver start-up energy inventory. Non- ycsd
t1 P ðyt1  yt Þ þ ðycsb
t1  t Þ
ycsb 8t 2 T : t P 2 ð7gÞ
negativity for xrt is ensured via Constraint (3c). Constraint (10c)
Constraint (5a) tracks start-up energy inventory, and Constraint
enforces binary requirements on the variables associated with gen-
(5b) allows nonzero inventory only during periods of cycle start-
erating usable thermal power, receiver start-up, receiver standby,
up. Constraint (5c) allows normal cycle operation only when
receiver shut down, and receiver start-up penalties.
start-up has been completed, when the cycle was previously oper-
ating, or when the cycle has been in standby mode. Constraint (5d)
2.4.2. Power cycle operations
limits the cycle start-up rate, and Constraint (5e) enforces a maxi-
Power cycle operation constraints largely mirror those of recei-
mum thermal power consumption limit by the power cycle. When
ver operations and include:
operating, the cycle must produce a minimum amount of power
Cycle Start-up
enforced by Constraint (5f). Constraint (6a) determines electrical
ucsu
t
c csu
t1 þ D  Q yt
6 ucsu 8t 2 T : t P 2 ð5aÞ power production based on a linear cycle performance curve and
ut 6 Myt 8t 2 T
csu csu
ð5bÞ the ambient temperature efficiency. The positive change in electri-
ucsu cal power production is determined by Constraint (6b). The pres-
yt 6 t c þ yt1 þ ycsbt1 8t 2 T : t P 2 ð5cÞ ence of w _ dt in the objective function provides a disincentive to
E
xt þ Q yt 6 Q yt 8t 2 T
c csu u
ð5dÞ vary power production from one time step to the next, thereby
xt 6 Q u y t 8 t 2 T ð5eÞ reducing system cycling and more closely representing operator-
preferred generation profiles. The appropriate magnitude of this
xt P Q l yt 8t 2 T ð5fÞ
penalty parameter is unknown but is explored further in a sensitiv-
Power Supply and Demand ity analysis provided in Section 4. Constraints (6c) and (6d) ensure
that if the net power production upper limit is greater than or
_t 6
gamb equal to that of the lower limit in any given time period, then that
w t
ðgp xt þ yt ðW u  gp Q u ÞÞ 8t 2 T ð6aÞ
gdes former production level must exceed that actually produced when
w_ dt P w _t w_ t1 8t 2 T : t P 2 ð6bÞ efficiency is accounted for, less that from parasitics due to pump-
if W _ net P W_ min then: ing power, heliostat field start-up, heliostat field tracking, power
t t
! cycle standby, and tower piping heat trace. If the net power pro-
_ rsb Ehs
W
_ net _ t ð1  gct Þ  Lr ðxrt þ xrsu
Pw c
t Þ  xt L  y t þ
rsu duction upper limit is less than the lower limit in any given time
W t
D D period, the production level is zero. Start-up mode persistence is
W _b
_ h yr  ycsb W 8t 2 T ð6cÞ prevented in Constraint (7a). Standby mode can persist according
t t
else: to the analogous receiver requirements (Constraint (7b)). Standby
_ t ¼ 0 8t 2 T
w ð6dÞ and start-up modes cannot coincide (Constraint (7c)), nor can
standby and power-producing mode (Constraint (7d)). Constraint
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 965

(7e) enforces the penalty logic for start-up from an off state while xrt ; xrsu
t ; ut ; ut
rsu csu
P 0 8t 2 T ð10aÞ
(7f) enforces the penalty logic for start-up from a standby state. _ dt ; st P 0; st 6 Eu 8t 2 T
xt ; w ð10bÞ
Constraint (7g) enforces the logic for shut-down from a power-
producing or standby state. Constraint (10b) ensures non- yrt ; yrsu rsup
t ; yt ; yrsb rhsp
t ; yt ; yrsd
t 2 f0; 1g 8t 2 T ð10cÞ
negativity for cycle start-up energy inventory, electrical power yt ; t ;
ycsu ycsb csup
t ; yt ; ychsp
t 2 f0; 1g 8t 2 T ð10dÞ
generation, and positive change in electricity production. Non-
negativity for xt is ensured via Constraint (5f). Constraint (10d)
enforces binary restrictions. 2.4.4. Cycle part-load correction
An optimized dispatch profile may result in electricity produc-
2.4.3. Energy balance tion lower than the CSP plant design-point during certain time
The energetic state of the system implicates positive and nega- periods in order to conserve stored thermal energy for more favor-
tive power terms, and the charge state of thermal storage (st ) able future market conditions, or to avoid penalties associated with
accounts for the cumulative difference between them. Several shut-down and start-up, for example. However, power cycle effi-
additional constraints regarding TES state of charge are enforced ciency is adversely affected by departure from design, as shown
as follows: in Fig. 2 [28].
The relationship between thermodynamic efficiency and ther-
st  st1 ¼ D  ½xrt  ðQ c ycsu b csb
t þ Q y t þ xt þ Q
rsb rsb
yt Þ mal input is nonlinear and, consequently, poses computational
8t 2 T : t P 2 ð8aÞ challenges. In order to improve tractability in the corresponding
st optimization model, an approximately linear function of cycle
xtþ1 þ Q b ycsb 6 rs  M  ð3 þ yrsu
tþ1 þ yt þ ytþ1 þ yt þ ytþ1 Þ
csb csb
thermal power consumption resolves the nonlinearity gcycle ðxt Þ  xt
tþ1
Dtþ1
by modeling electrical output, shown in Constraint (6a). The linear
8t 2 T : t 6 T  1 ð8bÞ
coefficient is the quotient of the difference between the minimum
Constraint (8a) ensures that energy into and out of TES balance with and maximum output from the power cycle and the corresponding
the charge state, and the conversion from power to energy intro- expression for the thermal power input.
duces a time step parameter D. Constraint (8b) addresses an artifact
Wu  Wl
arising from the difference between the modeling time resolution gp ¼ ð11Þ
(hourly) and the amount of time required to start the plant, which Qu  Ql
may not be in units of whole hours. If the power cycle is either run-
ning or in standby in time step t and in time step t þ 1, and if the 2.5. Dispatch model implementation
receiver starts up in time t þ 1, then the minimum charge level in
TES in time t þ 1 must be sufficient to carry operation through The typical model instance contains 912 variables and 1,615
the receiver start-up period. Note that yt þ ycsb
t 6 1 is enforced else- constraints. AMPL and CPLEX presolve reductions result in a prob-
where. Eq. (9) determines the expected fraction of each time step lem with an average of 442 variables and 652 constraints, and an
that would be used for receiver start-up, if applicable. average run time on a Dell PowerEdge R410 server running Ubuntu
8 8 99 14.04 with 12 GB RAM, 16 Intel processors at 2.72 GHz each of
< < Ec ==
Drs l
¼ min 1; max D ; n o ð9Þ 0.43 s per 48-h horizon evaluation. By contrast, implementation
t
: : max ; Q D ;;
in of the model using LPSolve [29], which is a freeware MIP solver
tþ1
platform for C++, requires an average of 0.83 s per solve. Presolve
Constraints (8a)-(8b) only track TES state of charge based on energy reductions are less effective, producing instances with 890 vari-
flow bookkeeping, not temperature. Accounting for energy quality ables and 920 constraints.
in the TES system via temperature of the molten salt introduces The number of time steps in the time horizon (T ) must be cho-
non-linear complexity and is not necessary in this formulation, as sen with care, as it greatly affects the typical model described here
previously discussed. as well as system techno-economic performance. The following
Variable bounds are enforced in (10a)–(10d), with (10b) bound- considerations are relevant when choosing a time horizon dura-
ing both the minimum and maximum amount of energy in storage. tion: (i) the problem complexity grows exponentially with the time
horizon length, and consequently, the amount of time needed for
an annual simulation will also grow significantly; (ii) the opti-

Fig. 2. Cycle efficiency as a function of input thermal power represented using a Fig. 3. The impact of time horizon length (hours) on annual energy production and
piece-wise linear function. PPA price.
966 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

Table 3 3. Model implementation

Case study plant design and control parameters.

Parameter Units Value Fig. 4 illustrates the dispatch optimization model within SAM
Gross electrical output MWe 115 whose interface provides both input and output display. The user
Cycle design efficiency % 41.2 selects the technology and financial model, then modifies the
Cycle design thermal input MWt 278.1 inputs to emulate their technology configuration of interest, after
Cycle maximum output MWe 120.75
which SAM simulates technical and financial performance by send-
Cycle minimum output MWe 28.75
Cycle start-up energy MWt-hr 57.5 ing information from the interface to the SAM Simulation Core.
Cycle start-up time hr 0.5 Therein lies the molten salt power tower (MSPT) technology model
Cycle standby consumption MWt 23 that contains a solar field design algorithm called SolarPILOT and
Receiver max. output (relative⁄) – 1.2 detailed calculators for determining weather data and the perfor-
Receiver min. output (relative) – 0.25
mance of the collector, receiver, power block, and TES subsystems.
Receiver start-up energy (relative) – 0.25
Receiver start-up time (relative) – 0.2 The MSPT model simulates annual production by evaluating
Receiver HTF temperature °C 574 performance over a sequence of hourly time steps, at each of which
Heat rejection technology – Air cooled the CSP controller determines the best operational mode given the
Heliostat size m2 144.4
conditions endogenous and exogenous to the system. The CSP sol-
Maximum receiver flux kW/m2 1,000
Hours of TES hr 1; . . . ; 18
ver ensures that all of the interconnected inputs and outputs
Solar multiple – 0:8; . . . ; 3 among the calculators agree with respect to the thermodynamic
⁄
state of the system. In summary, the architecture in Fig. 4 charac-
Relative to receiver thermal input design point.
terizes a molten salt power tower plant with storage, in which the
hour-by-hour plant operation protocol is determined using a 48-h
time horizon that rolls forward in 24-h increments.
mized dispatch profile maximizes revenue within the allotted time The Production Forecast Model determines expected future ther-
horizon, and an insufficiently long horizon emphasizes near-term mal energy generation of the solar field. While it is possible to
production at the expense of future, higher-value time periods; implement a variety of techniques for predicting electricity pricing,
(iii) an optimal profile may require thermal energy to be held in ambient temperature, and direct normal irradiance, this paper uses
storage overnight, and an insufficiently long time horizon (e.g., ‘‘perfect forecasting” in which the model generates expected per-
24 hours) will fail to account for next-day requirements; and, (iv) formance by reading ahead in the weather file. MSPT incorporates
given limitations on the number of branch-and-bound iterations the time series data from the weather and pricing databases corre-
and/or computation time per solve, an increased horizon length sponding to the horizon over which the model is solved.
raises the likelihood of adopting a suboptimal dispatch profile, The heliostat field concentrates power on the receiver (Q helio )
t
thereby negatively affecting expected plant performance. Fig. 3
according to the instantaneous optical efficiency (gsft ), direct nor-
shows the impact of the time horizon length on the annual energy
production and power purchase agreement (PPA) price (discussed mal irradiance (dt ), and mirror area (Asf ).
in the next section) for the reference plant defined in Table 3.
Q helio
t ¼ gsft dt Asf ð12Þ

Fig. 4. Information flow in the MSPT model. The MIP formulation is solved as a simultaneous set of equalities and inequalities, and the hourly solution profile is used by the
CSP controller to set target power production levels and operational states over the subsequent operational time horizon.
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 967

The total expected solar field production is the nonnegative differ- first-principles modeling of thermodynamics and heat transfer
ence of incident power on the receiver and convective (Q con
t
v
) and phenomena. The model’s engineering performance behavior is val-
radiative (Q rad idated and discussed in detail in [28,31].
t ) thermal losses:
h i The MIP Mathematical Formulation, when solved with an appro-
Q in helio
t ¼ max 0; Q t  Q con
t
v
 Q rad
t 8t 2 T ð13Þ priate algorithm, determines the performance and operation of the
plant using the forecast model and various operational constraints
The following reduced-order relationships model the expected (see Section 2).
technical performance of the collector and receiver, providing a The Pricing Model calculates the PPA price, which is the mini-
reasonably accurate approximation of expected field productivity. mum value at which a power producer should agree to sell elec-
The collector field model generates a lookup table containing opti- tricity in order to ensure that a specified internal rate of return is
cal efficiency as a function of sun position, and the CSP controller achieved. The PPA price is a useful surrogate for the profitability
supplies this information to the forecast model. The complexity of a project in that it accounts for the variability in electricity value
of modeling the receiver thermal loss via convection and radiation with time of day and time of year. As it is applied in SAM, the PPA
from the heated surface necessitates a simplified forecasting price is multiplied by the hour-by-hour TOD or ‘‘tariff” rate to
model: an area-weighted average molten salt temperature is given determine the value of electricity generated by the plant over time.
as the weighted average of the inlet and outlet temperatures, SAM calculates the PPA price assuming a target internal rate of
where the coefficient is receiver-specific [30]. return (11% in the current study) and an annual escalation rate
of 1%. For this reason – and somewhat counter-intuitively – a
T eff ¼ 0:55  ðT out þ T in Þ ð14Þ low PPA price is desirable. From the perspective of a power pro-
Radiative losses are calculated at each time t as: ducer, a low PPA price improves its competitiveness. Alternatively,
  the PPA price could be specified and the internal rate of return
4 4
Q rad
t ¼ Arec r ðT eff Þ  ðT amb
t Þ ð15Þ maximized, and results from either approach would be equivalent.
The results here translate the objective function value of (R) into
in which Arec is the receiver surface area, r is the Stefan-Boltzmann PPA price by taking fixed costs as sunk and maximizing revenue
constant,  is the temperature-weighted surface emittance, and T amb
t
generated from electricity sales.
is the expected ambient dry-bulb temperature. Convective losses
4. Case studies
are expressed as a function of wind velocity for the molten salt
technology, scaled by radiative loss. The coefficients in (16) are
This study explores a range of plant TES sizes and solar multiples,
determined by regressing simulated data points that are generated
the latter of which is defined as the ratio of solar field thermal power
using the MSPT detailed receiver model.
  output to power cycle thermal input at design conditions. As the
Q con
t
v
¼ 5:645  104 V 3t þ 0:01561V 2t  0:00911V t þ 0:48124 Q rad
t
solar multiple increases, so too does the optimal amount of TES
and the resulting plant capacity factor, but these values may be cho-
ð16Þ sen independently. Table 3 provides a summary of key design
where V t is the wind velocity at time t. parameters which are obtained from the default SAM-MSPT case.
The Engineering Performance Model (consisting of the CSP con- For this analysis, SAM automatically determines the heliostat field
troller, CSP solver, and detailed performance calculators in Fig. 4) layout given the specified solar multiple and other design parame-
predicts plant behavior and productivity over time using computa- ters. Each evaluation takes as fixed the TES and solar multiple and
tionally expensive procedures derived from physically based, determines the optimal dispatch schedule for that system
configuration.
In addition, this analysis considers four market scenarios
(Fig. 5), three of which have been adopted from [22], and one of
which is the ‘‘generic summer peak” scenario used as the default
for the MSPT model. The two-tier tariff market encourages daytime
production with an evening spike. The pool price tariff introduces
an additional morning spike and weights incentives seasonally.
The fixed daytime tariff allows sales during daytime hours, but is
unique in its binary nature; no revenue is available during night-
time operation. Finally, the SAM generic peak schedule combines
features from the two-tier and pool price tariffs.
Using MSPT, this paper compares the dispatch optimization
methodology to the previous approach that relies on heuristic con-
trol which was configured to allow power generation any time the
TES state of charge exceeded the threshold for minimum power
cycle operation (satisfying Constraint (5f)). The cycle generates
power at the design-point level unless insufficient energy is avail-
able in storage. Power cycle start-up occurs whenever energy in
storage exceeds the quantity needed to deliver the start-up power
for a single time period. The heuristic allows power generation until
energy storage is exhausted each night, if applicable. This approach
emphasizes maximum energy generation throughout the year.

4.1. Results
Fig. 5. Market pricing scenarios presented by Guédez et al. [22]. These tariff
schedules are implemented to determine the impact of dispatch optimization on Table 4 presents the results of the PPA analysis, where the
system sizing. reported values correspond to the configuration with the
968 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

Table 4 An important feature of dispatch optimization is the apparent

Characteristics for each market scenario in which PPA price is at a minimum value, improvement in the consistency of production during high-value
both for heuristic (H) and optimized (O) dispatch.
time periods. Fig. 6 illustrates this behavior for the pool price tariff
Market scenario Solar mult. Hours TES PPA price scenario. Figs. 6a and b show the hourly TES charge state profiles
H O H O H O for each day of the year for both heuristic and optimized dispatch,
Two-tier 2.2 2.0 5 7 8.97 7.71
respectively. Also shown on the plot are the tariff multiplier
Pool price 2.2 2.2 8 9 9.66 8.75 schedules for summer (black) and winter (green) that determine
Fixed daytime 1.8 2.0 4 6 13.30 12.75 the revenue associated with generation during a particular hour
SAM default 2.6 2.6 12 14 10.41 10.09 of the day.
Dispatch optimization substantially changes the daily opera-
tional profile. Whereas heuristic dispatch allows TES to discharge
minimum PPA price for the indicated scenario. Dispatch optimiza- in the evening and late-night hours, optimized dispatch typically
tion successfully reduces PPA price compared to the heuristic reserves some quantity of TES to allow morning start-up. The TES
dispatch method. Heavily weighted schedules (i.e., pool price and profiles show that heuristic dispatch is much more operationally
two-tier) lead to more substantial PPA price reductions (about repetitive relative to optimized dispatch, implying that the latter
10–15%), indicating that dispatch optimization is an essential strategy accounts for expected resource availability and future
aspect of plant operation for ‘‘peaker” markets that provide rela- pricing when scheduling TES usage.
tively short time windows of high-value energy pricing. The reader Figs. 6c and d show the distribution of electricity generation for
can also observe that systems operating in markets with more each hour of the day over the course of the year. Each box-whisker
uniform tariff factors still benefit significantly from dispatch opti- plot describes the variability in electricity generation for each day
mization, which alters the size of TES and the solar multiple at at the specified hour, and the box limits indicate the extents of the
which PPA price is minimized. This implies that dispatch optimiza- first and third quartiles. The whiskers correspond to twice the
tion should not be relegated to operational analyses, but rather inner quartile range. Points that lie outside of this limit are plotted
should be part of the project screening and design process. individually and are distribution outliers.

(a) Daily thermal storage charge state profile - heuristic dispatch (b) Daily thermal storage charge state profile - optimized dispatch

(c) Annual variability of electricity generation - heuristic dispatch (d) Annual variability of electricity generation - optimized dispatch

Fig. 6. Comparison of performance profiles for the pool price tariff schedule. Plots (a) and (b) show traces of the TES charge state for each day of the year. Plots (c) and (d)
show box-whisker plots of daily electricity production variability over a year grouped by hour of the day. Each box indicates the mean annual electricity generation by hour,
the first and third quartile limits (box limits), and two times the interquartile range (whiskers). ‘‘Outliers” are shown as black dots. Summer (black) and winter (green) tariff
multipliers are overlaid on each plot. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 969

maximizing generation during periods of peak revenue and cycle

efficiency to improve output stability. If increased to an extreme
_
(C dW ¼ 10:0), the dispatch profile approaches that of a baseload
plant, only peaking for very short periods of time.
This analysis leads to several considerations regarding the opti-
mal design of a CSP plant given the potential ramping costs of the
cycle. Namely, if production change costs truly are on the order of
10$=DkWe, then the solar multiple, thermal storage, and power
cycle sizes should generally be chosen to mimic a baseload plant.
Penalties for frequent production changes would only be offset
for the case in which TOD pricing variations are very large. Further-
more, an operator might reduce dispatch profile complexity but at
some expense to apparent near-term plant profitability.
This paper also considers two TOD pricing scenarios (‘‘Default”
and ‘‘Peaker”) in which the hours of TES and solar multiple are
Fig. 7. Plant power generation profile with varying change in production penalty optimized for the no-penalty case. The first scenario is the generic
_
values, C dW . (Penalties are given in the legend.)
summer peak profile in SAM, and the second is the pool price pro-
file with spikes in pricing in the morning and evening and a price
Fig. 6c shows that heuristic electricity production is highly vari-
reduction during the middle of the day.
able in the early morning, evening, and nighttime hours (tall
Fig. 8 shows the impact of increasing the production change
boxes) and is less variable during daytime hours (short boxes).
cost penalty on annual energy generation, PPA price, and number
The variability in electricity generation is inversely related to solar
of turbine cycles per day, which is calculated as the total annual
resource availability – an intuitive observation. However, produc-
positive change in cycle production normalized to a value that cor-
tion is highly variable during peak revenue hours as shown by
responds to ramping the cycle from off to full load once a day for
the tariff multipliers. Fig. 6d depicts optimized dispatch in which
the entire year. If the number of cycles per day is greater than
electricity generation variability is reduced during peak revenue
unity, then the turbine experiences more than one full cycle per
hours.
day on average.
Several interesting observations arise from this analysis. First,
4.2. Penalty parameter sensitivity study the number of cycles per day decreases as the production change
cost penalty increases, as intended. If production variability is
This study also seeks to understand the impact of operational not penalized at all, the optimal solution results in more than
cost parameters on both the generation profile and on PPA price. one cycle on average per day for both the Default and Peaker cases.
Plant operation that requires multiple daily start-ups or frequent As the penalty for changing production increases, the number of
production ramping may be difficult to execute and lead to addi- cycles drops significantly, which may be an important factor in
tional maintenance costs over time. This paper quantifies the rev- increasing plant lifetime and reducing maintenance costs. Remark-
enue and production impact of obtaining ‘‘desirable” operational ably, the number of cycles can be reduced by 50% or more by
profiles (that is, relatively consistent output with few starts or manipulating this penalty without significant degradation of the
ramp events). objective function value.
_
The production change penalty, C dW , penalizes any positive Second, the trends in reduction of cycles per day, annual energy
change in power cycle electricity production from one hour to generation, and PPA price are mirrored between the Default and
the next. Fig. 7 presents the dispatch profile over several selected Peaker cases. Both show that production change cost in the range
days in September in which four different penalty values are used. of 0.5–2.0 $/kWe minimizes PPA price and number of cycles per
_ day, though the true costs of turbine ramping are not known and
As the value of C dW increases, the optimal solution sacrifices
thus omitted from the PPA calculation. Non-coincidence of this

Fig. 8. Impact of production change cost penalty on the number of turbine cycles
per day, annual energy generation, and PPA price for two pricing scenarios – a
generic summer afternoon peak schedule (Default) and a morning/evening double-
peak (Peaker) schedule. Annual energy and PPA price are shown as fractional values Fig. 9. Number of cycle starts per year, annual energy output, and PPA price for the
relative to the lowest-penalty case. Default case with varying scenarios for cycle start-up cost.
970 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971

outcome may imply an important cost threshold regarding CSP In summary, this model provides a methodology to optimize
plants in general. Lastly, increasing the penalty leads to reduced the trade-offs between CSP component and subsystem perfor-
annual output and increased PPA price below a threshold corre- mance, the effects of demand, and the amount of revenue obtained
sponding to 0.4–0.6 cycles per day. under various market schedules. Future work will incorporate,
A final study considers the cycle start-up cost penalty’s impact inter alia, forecast uncertainty, and more precise cost estimates
on the same performance metrics. This penalty is related to ramp- of component operations and maintenance requirements. It will
ing cost, but differs in that it represents a penalty incurred for a also examine how plant design and maintenance affect the overall
discrete event that occurs only when the power cycle transitions cost and nature of the dispatch strategy.
from an off state to an on state, after which this penalty does not
influence operation. Fig. 9 shows the result of varying start-up cost Acknowledgments
for the Default case.
As with the production change penalty, the start-up penalty can This work was funded by the United States Department of
significantly affect the behavior of the power cycle. A small cost of Energy – Energy Efficiency and Renewable Energy under award
$100 per start leads to a relatively large annual number of cycle numbers DE-EE00025831 and DE-EE00030338. The authors grate-
starts (about 250). As a point of comparison, the number of cycle fully acknowledge Jennifer DiCarlo at Colorado School of Mines for
starts incurred using heuristic dispatch is 370 per year. The num- her contributions to the MIP mathematical formulation, Mark
ber of starts remains fairly constant (within variance that is to be Mehos at the National Renewable Energy Laboratory for guidance
expected based on numerical error in the solution) until the cost on market factors, and Charles Diep and Jolyon Dent at SolarRe-
increases by a factor of 100, and a factor increase of 1,000 reduces serveÒ for feedback on modeling priorities and plant operations.
starts by approximately 50% without a significant effect on annual
energy output and PPA price. Therefore, operational protocols that
References
seek to minimize full cycle starts and stops can theoretically offer
equally viable financial performance compared to more traditional [1] Margolis R, Coggeshall C, Zuboy J. Sunshot vision study. US Dept. of Energy,
approaches. February.
[2] Gil A, Medrano M, Martorell I, Lázaro A, Dolado P, Zalba B, Cabeza LF. State of
the art on high temperature thermal energy storage for power generation. Part
4.3. Applications 1-Concepts, materials and modellization. Renew Sustain Energy Rev 2010;14
(1):31–55.
[3] Jorgenson J, Denholm P, Mehos M. Estimating the value of utility- scale solar
The results shown herein are readily applicable in practice, both
technologies in California under a 40% renewable portfolio standard. Tech. rep.
for modeling and plant operations applications. First, modeling May. Golden, CO: National Renewable Energy Laboratory; 2014.
activities are fundamental to research, project development, and [4] Turchi C, Mehos M, Ho CK, Kolb GJ. Current and future costs for parabolic
trough and power tower systems in the US market preprint. Renew Energy
policymaking decisions, and an accurate estimate of technology
2010(October):11.
performance directly impacts each of these areas. Researchers [5] Dunn RI, Hearps PJ, Wright MN. Molten-salt power towers: newly commercial
require tools that quantify the impact of advances in the technol- concentrating solar storage. Proc IEEE 2012;100(2):504–15.
ogy, and utilize models to identify research priorities. The provi- [6] Burgaleta JI, Arias S, Ramirez D. Gemasolar: the first tower thermosolar
commercial plant with molten salt storage system. In: SolarPACES conference.
sion of a dispatch optimization tool in a publicly available [7] Herrmann U, Kelly B, Price H. Two-tank molten salt storage for parabolic
software package enables assessment of the value of new thermal trough solar power plants. Energy 2004;29(5–6):883–93.
energy storage technologies – e.g., [32], or power cycles – e.g., [33], [8] Denholm P, Mehos M. Enabling greater penetration of solar power via the use
of CSP with thermal energy storage. Tech. rep. November, Golden, CO; 2011.
that otherwise may not interact with other subsystems as antici- [9] Pruitt KA, Braun RJ, Newman AM. Evaluating shortfalls in mixed-integer
pated upon deployment. Project developers rely on models for ini- programming approaches for the optimal design and dispatch of distributed
tial plant design, attaining financing, project permitting, and generation systems. Appl Energy 2013;102:386–98.
[10] Blair N, Dobos AP, Freeman J, Neises T, Wagner M, Ferguson T, Gilman P.
finally, during plant operation. Dispatch optimization tools such System advisor model, SAM 2014.1.14: general description. Tech. rep.
as this are valuable for these purposes, and the authors present February, National Renewable Energy Laboratory, Golden, CO; 2014.
additional applied results and model validation in [34]. [11] National Renewable Energy Laboratory, System Advisor Model; 2012.
[12] Perera A, Attalage R, Perera K, Dassanayake V. A hybrid tool to combine multi-
objective optimization and multi-criterion decision making in designing
standalone hybrid energy systems. Appl Energy 2013;107:412–25.
5. Conclusions [13] Nosrat A, Pearce JM. Dispatch strategy and model for hybrid photovoltaic and
trigeneration power systems. Appl Energy 2011;88(9):3270–6.
This paper develops and implements a mixed-integer program- [14] Pandžić H, Kuzle I, Capuder T. Virtual power plant mid-term dispatch
optimization. Appl Energy 2013;101:134–41.
ming model within SAM to optimize the TES dispatch schedule for [15] Martinez-Rojas M, Sumper A, Gomis-Bellmunt O, Sudrià-Andreu A. Reactive
a molten salt power tower plant; this schedule provides a target power dispatch in wind farms using particle swarm optimization technique
power generation profile that is used in conjunction with a simu- and feasible solutions search. Appl Energy 2011;88(12):4678–86.
[16] Thorin E, Brand H, Weber C. Long-term optimization of cogeneration systems
lation model that evaluates plant performance at an hourly level in a competitive market environment. Appl Energy 2005;81(2):152–69.
over a year-long time horizon. SAM’s detailed performance model [17] Cho H, Mago PJ, Luck R, Chamra LM. Evaluation of CCHP systems performance
mitigates some of the approximations present in the MIP based on operational cost, primary energy consumption, and carbon dioxide
emission by utilizing an optimal operation scheme. Appl Energy 2009;86
formulation. (12):2540–9.
The results indicate that dispatch optimization can significantly [18] Fürsch M, Hagspiel S, Jägemann C, Nagl S, Lindenberger D, Tröster E. The role of
improve plant revenue, though the gains vary with plant capacity grid extensions in a cost-efficient transformation of the European electricity
system until 2050. Appl Energy 2013;104:642–52.
factor and electricity markets. In particular, scenarios with heavily [19] Parisio A, Rikos E, Tzamalis G, Glielmo L. Use of model predictive control for
weighted pricing schemes or narrow windows of high revenue experimental microgrid optimization. Appl Energy 2014;115:37–46.
benefit the most, e.g., PPA price (indicative of the profitability of [20] Zheng Y-J, Chen S-Y, Lin Y, Wang W-L. Bio-inspired optimization of sustainable
energy systems: a review. Math Problems Eng 2013.
the plant), can improve by 10–15%. Plant revenue is negatively
[21] Hirsch T, Blanco MJ, Feldhoff JF, Eck M, Wagner MJ. Standardization of CSP
affected by the energetic and financial cost of starting the solar performance model projection latest results from the stamp project. In:
receiver and power cycle equipment, and this paper shows that Proceedings of the ASME 2011 international conference on energy
optimized generation profiles can achieve a reduction in the num- sustainability. p. 1–6.
[22] Guédez R, Ferragut F, Défense PL, Topel M, Callaba I, Pérez-Segarra CD. A
ber of turbine starts per year by 50% or more – in some circum- methodology for determining optimum solar tower plant configurations and
stances – with little impact on project financial performance. operating strategies to maximize profits. In: Proceedings of the ASME 2015 9th
 M.J. Wagner et al. / Applied Energy 203 (2017) 959–971 971

international conference on energy sustainability. San Diego, CA: American [29] Berkelaar M, Eikland K, Notebaert P. LPSolve: open source (mixed-integer)
Society of Mechanical Engineers. p. 1–15. linear programming system. Eindhoven U. of Technology.
[23] Fourer R, Gay D, Kernighan B. AMPL, vol. 119. Boyd & Fraser; 1993. [30] Lata JM, Rodríguez M, de Lara MÁ. High flux central receivers of molten salts
[24] IBM. IBM ILOG CPLEX V12.1 user’s manual for CPLEX; 2009. for the new generation of commercial stand-alone solar power plants. J Sol
[25] Madaeni SH, Sioshansi R, Denholm P. How thermal energy storage enhances Energy Eng 2008;130(2):021002.
the economic viability of concentrating solar power. Proc IEEE 2012;10 [31] Wagner MJ, Gilman P. Technical manual for the SAM physical trough model,
(2):335–47. Tech. rep. June. Golden, CO: National Renewable Energy Laboratory; 2011.
[26] Sioshansi R. The value of concentrating solar power and thermal energy [32] Singh D, Zhao W, Yu W, France DM, Kim T. Analysis of a graphite foam-NaCl
storage. IEEE Trans Energy 2010;1(3):173–83. latent heat storage system for supercritical CO2 power cycles for concentrated
[27] Guédez R, Spelling J, Laumert B, Fransson T. Reducing the number of turbine solar power. Sol Energy 2015;118:232–42.
starts in concentrating solar power plants through the integration of thermal [33] Neises T, Turchi C. A comparison of supercritical carbon dioxide power cycle
energy storage. In: Proceedings of the ASME turbo expo 2013, San Antonio, configurations with an emphasis on CSP applications. Energy Proc
Texas. p. 1–10. 2014;49:1187–96.
[28] Wagner MJ. Simulation and predictive performance modeling of utility-scale [34] Wagner MJ, Newman A, Braun R, Dent J, Diep C. Dispatching power at a
central receiver system power plants [Masters thesis]. Madison: University of concentrated solar energy facility. Working paper.
Wisconsin; 2008.