Accepted Manuscript Optimization of a stand-alone photovoltaic–wind–diesel–battery system with multilayered demand scheduling Tu Tu, Gobinath P. Rajarathnam, Anthony M. Vassallo PII:
S0960-1481(18)30826-7
DOI:
10.1016/j.renene.2018.07.029
Reference:
RENE 10303
To appear in:
Renewable Energy
Received Date: 11 December 2017 Revised Date:
6 June 2018
Accepted Date: 7 July 2018
Please cite this article as: Tu T, Rajarathnam GP, Vassallo AM, Optimization of a stand-alone photovoltaic–wind–diesel–battery system with multi-layered demand scheduling, Renewable Energy (2018), doi: 10.1016/j.renene.2018.07.029. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Optimization of a Stand-Alone
2
Photovoltaic–Wind–Diesel–Battery
3
System with Multi-Layered Demand
4
Scheduling
5
Tu Tu, Gobinath P. Rajarathnam, and Anthony M. Vassallo*
6
The University of Sydney, School of Chemical and Biomolecular Engineering, Sydney, NSW
7
2006, Australia.
M AN U
SC
RI PT
1
8
*Corresponding author contact information
9
Corresponding author: Anthony M. Vassallo Telephone: +61 2 9351 6740
11
Fax: +61 2 9351 2854
12
Email:
[email protected]
13
Postal address: School of Chemical and Biomolecular Engineering, J01, University of Sydney,
14
NSW 2006, Australia
15
E-mail addresses of other authors
16
Tu Tu:
[email protected]
17
Gobinath P. Rajarathnam:
[email protected]
18
Version: 0.10b
AC C
EP
TE D
10
1
ACCEPTED MANUSCRIPT
1. Abstract
20
Operational and financial optimization of a renewable energy-based stand-alone electricity
21
micro-grid is described. Due to the large problem size in time-series models, we construct
22
the model using mixed integer linear programming (MILP). As the constraints required in this
23
model generally have modest complexity, we were able to perform piece-wise linearization
24
on any non-linear variable relationship. Additionally, controls have also be applied on the
25
demand side. Here, a two stage MILP model has been developed to minimize the overall
26
levelized electricity cost for a micro-grid containing a photovoltaic power source, wind
27
turbine, diesel generator, and an energy storage system. The model aimed to converge on a
28
balance of decision accuracy and computational efficiency. Model outputs were capable of
29
defining both the optimal system sizing and scheduling for each system component, with
30
additional demand management control levers on the loss of power supply probability and
31
load deferring allowance. We believe that this model is one of the first to explore the
32
possibilities of the influences of potential demand management strategies in overall system
33
cost reduction, while presenting a relatively efficient first-pass component sizing for stand-
34
alone micro-grids.
35
Keywords
36
MILP; demand scheduling; load shifting; off-grid; optimization; energy storage
SC
M AN U
TE D
EP AC C
37
RI PT
19
2
ACCEPTED MANUSCRIPT
2. Introduction
39
In recent years, both renewable and alternative energy sources which generate power
40
discretely from naturally replenishing resources are getting more attention, especially in
41
rural locations. Renewable energy systems operate with high reliability and low
42
maintenance requirements, and emits less greenhouse gas (GHG) compared to fossil fuel-
43
based combustion energy sources. Therefore renewable energy sources are often suitable
44
for remote and inaccessible areas. In the context of islands without local fossil fuel
45
production, fossil fuel based generators are operationally expensive due to the high
46
transportation costs of the fuel to site as well as the fuel cost. Therefore it is reasonable to
47
consider a renewable energy source driven stand-alone micro-grid for islands and other
48
isolated regions. This would provide competitive energy costs while reducing the energy
49
dependence of the island to the mainland, thereby increasing the island’s self-sustainability.
50
Apart from generating low carbon electricity, demand side management could also be used
51
in micro-grids to improve the balance between supply and demand. By enabling some
52
demand management strategies, the burden placed on battery storage units were
53
significantly reduced, which resulted in a reduced overall system cost, hence a more
54
affordable levelized energy cost.
55
De Groot et al. [1] analysed the impact of demand response in isolated power systems,
56
concluding that recent technical studies have followed a trend of focusing on the supply side
57
(namely renewable energy systems), with little consideration of demand management
58
possibilities. The authors noted that consideration of the latter could be a promising way to
59
mitigate demand–supply balancing problems caused by intermittent energy supply, hence
60
directly influencing component sizing optimization and operational improvements, and
61
finally translating into an overall system cost reduction. Fezai et al. [2] proposed a study on
62
load amplitude clipping and load shifting for a stand-alone photovoltaic system, concluding
63
that acting on the load profile to match generation is a positive option toward higher
64
performances. Bilal et al. [3] modeled the levelized cost of energy on three independent load
65
profiles, whereas the demand profile correlated the best with PV generation profile and
66
yielded the lowest levelized cost of energy.
67
Following from that point, stand-alone micro-grid applications are suitable candidates for
68
renewable energy-centric system designs. Wang et al. [4] investigated and simulated the
69
operational performance of a solar-wind-fuel cell based energy system for stand-alone
70
applications. Yang et al. [5] proposed a sizing method for stand-alone hybrid solar-wind
AC C
EP
TE D
M AN U
SC
RI PT
38
3
ACCEPTED MANUSCRIPT systems, utilizing five objective variables: number of PV/WT/Bat modules, PV tilt angle and
72
turbine installation height. The authors discussed impacts of loss of power supply probability
73
(LPSP) on the levelized electricity cost (LEC), i.e. the fitness variable of the model. It was clear
74
that as maximum LPSP increased, LEC decreased with it. The impact was most significant
75
during low LPSP allowances, which was potentially more relevant to practical applications as
76
high LPSP could result in severe interruptions to services and operations. Going from 1% to 2%
77
LPSP, the PV module and wind turbine power were more moderate, yielding an levelized
78
system cost to be 8.5% lower, thereby indicating the sensitivity of the output to the LPSP
79
parameter.
80
Merei et al. [6] modeled an off-grid hybrid photovoltaic-wind-diesel system with different
81
battery technologies to power a constant AC load. The study incorporated three battery
82
technologies in order to explore the opportunities in constructing an effective battery power
83
bank that consists of multiple battery types, to maximize each battery technology’s
84
strengths and minimize its weaknesses. The authors emphasized system reliability and
85
availability, including a diesel generator which was sufficient to supply the load by itself was
86
enforced into the designed system, hence ensuring no loss of power supply. Another
87
emphasis was the possibility of different battery technologies working together in the same
88
system. The authors discussed and modeled in detail three types of batteries (lead-acid,
89
lithium-ion and vanadium redox-flow), with the important characteristics of each battery
90
type formulated and quantified, such as the battery state-of-health, cycling, and ageing
91
effects. Thresholds were placed on each battery, such as component capacity degradation
92
and replacement. The model used net present value as the objective and optimized to
93
minimize the overall system cost, under the condition of no loss of supply, and with the
94
freedom of utilizing one or more batteries of arbitrary size. The optimization result
95
suggested that the renewable energy system set up was cost effective compared to the
96
diesel only set-up, by having energy cost about 50% cheaper. However, multiple battery
97
configurations were not favoured against single battery technology configurations utilizing
98
only the redox-flow battery, due to two potential reasons: the strengths of lithium-ion
99
batteries in high energy density and high cell voltage were not required for the project, and
100
the relatively short lifespan in lead acid battery translated to high maintenance costs and
101
was not favoured. Vanadium redox-flow battery, despite low energy density and
102
charge/discharge efficiency, the long lifespan and low cost made it to be the most suitable
103
battery technology for the project. This finding suggests that Li-ion is not a unique solution
104
to the issue of widespread energy storage.
AC C
EP
TE D
M AN U
SC
RI PT
71
4
ACCEPTED MANUSCRIPT Shang et al. [7] discussed an unconventional amenity metric to describe the quality of power
106
supply: instead of LPSP (discussed in most of the open literature), the authors quantified and
107
correlated each end-use demand to comfort level, including air quality or carbon dioxide
108
level from ventilation systems, internal temperature from air-conditioning systems and
109
illumination from lighting. This relatively unique approach differs from allowing a maximum
110
LPSP over the modeling period, as the mathematical model set a discomfort level
111
representing the combined occupant discomforts caused by the unfulfilled end-use demands.
112
This approach is an improvement over the LPSP approach in two ways: firstly, it was applied
113
to each time-series point in the modeling, compared to the LPSP approach that was placed
114
on the whole model. The comfort level approach allowed better visualization on the
115
modeled result on unfulfilled time points, because it is highly likely that the LPSP approach
116
would completely sacrifice the amenity of the scheduled loss of power points, in order to
117
maintain resources for better optimization of the objective. Secondly, the discomfort
118
approach allowed discrete settings of discomfort for each time-series point. An array of
119
discomfort constraints could be imported into the model to better reflect the occupants
120
comfort level desires. Although similar implementations could be made for the LPSP
121
approach, it lacked flexibility as each time-series points could only be set with a Boolean
122
variable to either allow or disallow loss of electricity. However, the addition of comfort level
123
quantification introduced new variables and constraints in the optimization model, making it
124
more complicated and harder to solve. At the same time, the weighting of each end-use to
125
comfort level could be highly subjective, hence the weighting variables had to be selected
126
carefully to reflect the reality.
127
Genetic algorithms (GA) have been increasingly used for such computations, such as in the
128
work by Koutroulis et al. [8], where the authors proposed a detailed model of a hybrid
129
photovoltaic–wind–battery (PV-WT-BA) system. Their proposed model included modular
130
system specifications and was aimed at optimizing over the system’s lifetime, accounting for
131
capital and maintenance costs. The objective function used covered aspects such as the
132
number of components required in each of the modules, photovoltaics tilt angle, and the
133
wind turbine installation height. A two staged GA approach was utilized, where the first
134
optimization stage focused on system configuration on a general scale, while the second
135
stage took the output of the first system and generated another computational efficient
136
model using the known information, seeking global maximum values for a finer resolution of
137
output result.
AC C
EP
TE D
M AN U
SC
RI PT
105
5
ACCEPTED MANUSCRIPT Nafeh [9] proposed a simplified alternative to the PV-WT-BA system by Koutroulis et al., with
139
non-integer variables. While the fitness function of the former and latter was the same (to
140
minimize the total system cost), the latter was constructed as a theoretical model with the
141
pricing section utilizing units such as unit cost of PV panels, normalized WT swept area, and
142
battery capacity. Additionally, PV generation was set to be a linearly proportional to solar
143
irradiance, while WT generation was set to be cubically proportional to the wind speed
144
(typically a value between the cut-in speed and rated wind speed). However, while these
145
simplifications in model construction would reduce computational runtime, it is noted that
146
there could be increased inaccuracies in cost estimation. While PV panels and battery
147
systems are available in relatively smaller modules, WT towers are larger in terms of both
148
capacity and physical dimensions (and hence better on a cost-basis to be modeled
149
modularly). Modular pricing provides a more realistic indication of system sizing and costs,
150
as well as limits the stepwise WT sizing by putting constraints on the GA solver and forcing it
151
to find optimal solutions which utilize integer numbers of WTs. This approach provides an
152
improved estimation of the costs involved if the modeled project was to be commercially
153
realized. Consequently, a similar approach to modeling is adopted in the present work, since
154
limiting the system component sizing as modular is computationally efficient while
155
maintaining a reasonable accuracy of the final obtained results (in terms of costs and sizing).
156
3. Methodology
SC
M AN U
3.1.
TE D
157
RI PT
138
Modeling concept
The hybrid stand-alone micro grid consists of three sources of electricity generation, one
159
electrical load and one energy storage device, all connected via an AC bus and a DC bus, and
160
communicated through a bi-directional inverter/rectifier. An additional load dump was
161
added to manage any excess load when storage was full and generation exceeded demand.
162
Solar and wind were designed to be the primary source of energy generation, whereas diesel
163
generator was installed as a boost generator to smooth the variable nature of renewable
164
generation. Battery storage was also installed to reduce excess electricity and to assist other
165
generation sources in adverse supply-demand conditions. The load controller would
166
calculate the optimal scheduling of both battery storage system and demand management,
167
to maximize the effectiveness of renewable generation and minimize cost. The time-series
168
electricity demand data was a synthetized representation of costal islands near Sydney area,
169
generated from the local electricity consumption and usage behavior. TMY (Typical
170
Meteorological Year) data was selected for weather related inputs. We were aware that the
AC C
EP
158
6
ACCEPTED MANUSCRIPT fluctuations in fine resolution data would cause results to fluctuate and potentially affect the
172
final outcome, we felt TMY data was still the most representative of the local climate and
173
most suitable for this study, as it focuses on the feasibility analysis instead of sensitivity test.
174
Figure 1 presents an illustration of the connectivity of the various system components
175
modeled in the present work.
M AN U
SC
RI PT
171
176
TE D
179
3.2. Mathematical modeling and optimization of stand-alone hybrid systems
EP
178
Figure 1. Illustration of the problem being modeled in the present work.
AC C
177
7
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
180
Figure 2. Schematic of the two-staged optimization model used in the present work.
182
The formulation of the problem consists of two nested mixed-integer linear programming
183
(MILP) optimization models, adopting similar concept but to fulfil different purposes. Due to
184
the nature of MILP formulation, decision variables are only allowed to have linear
185
relationships with other variables. Unfortunately most system components have non-
186
linearity in their operational profile, so to accurately model these behaviors, piece-wise
187
linearization techniques were utilized for several components in the system, mainly
188
generator efficiency and battery charging/discharging energy loss. Due to the uncertainties
189
in component sizing, piece-wise linearization had to be dynamically performed for every
190
time-series point on every component sizing alternation. The computational complexity was
191
increased by the large number of integer and Boolean variables required to perform piece-
192
wise linearization. With hourly time-series resolution and modeling length being a whole
193
year, the amount of slack variables generated during problem solving not only required very
194
large amount of system memory, it also greatly increased the run time. To populate a multi-
195
dimensional matrix consisting of permutations of different load deferring and loss of
196
electricity parameters, the model had to be run many times under different settings. On low
197
tolerance gap allowance setting, solving every point in the matrix could be highly time
198
consuming and unpractical.
AC C
EP
TE D
181
8
ACCEPTED MANUSCRIPT MILP has been the algorithm of choice in previous publications, such as studies by Mazzola
200
et al. [10] focused on optimal scheduling of units in an energy management system over a
201
particular time horizon, and by Chen et al. [11] to quantitatively examine the optimal size for
202
both grid-connected and isolated micro-grids. Morais et al. [12] proposed a constrained
203
MILP model for a time-series problem on optimal scheduling of a renewable micro-grid in an
204
isolated load area, with pre-determined component sizing and evaluation of the generation
205
cost. The model objective function was formulated to minimize the overall generation cost
206
over the modeled interval. Dai et al. [13] formulated a model featuring electricity demands
207
being pre-determined then managed by a controller to make best use of the electricity
208
generation from renewable sources. Depending on the energy supply condition, the load
209
would be either fulfilled or curtailed. A priority load control algorithm was developed by
210
Faxas-Guzmán et al.[14], in which demand was tagged with different priority levels, and
211
during adverse generation conditions, loads with higher priority would be fulfilled first.
212
The concept of load deferring has been considered in other researches. Bekele et al. [15]
213
[16]used the software tool Hybrid Optimization Model for Electric Renewables (HOMER) to
214
perform feasibility study for renewable energy focused system in Ethiopia. Deferrable load
215
in the form of water pumping was considered.
216
A two-stage configuration was proposed to resolve this issue and speed up the process,
217
without sacrificing model accuracy. The model was split into the initial sizing and feasibility
218
stage, and then followed by the operational scheduling stage. First stage was constructed to
219
run on a much shorter time, effectively solving a much smaller problem. In addition, a large
220
relative gap was implemented for the first stage modeling, meaning that the termination
221
condition of the solving was relaxed, as an optimizer solver will attempt to approach the
222
solution from both sides, finding improvements to the current best solution, and estimating
223
the currently unknown global optimal solution. By allowing a larger gap between these two
224
points, the solving was sped up significantly because for NP hard mix integer problems, the
225
final convergence could often be computationally heavy. With the problem size and
226
termination condition relaxed, the first stage modeling worked out the general direction of a
227
global optimal solution, providing information on
228
information was then fed to the second stage problem, so it could utilize better variable
229
bound ranges and search spaces to solve a bigger problem. Finally an iterative approach was
230
conducted as a search for global optimum solution, iterating on permutations of different
231
component sizes based on the estimation given by stage one. As a result, stage two of the
232
optimization process would calculate the optimal system scheduling and running costs, and
AC C
EP
TE D
M AN U
SC
RI PT
199
decision variables estimations. This
9
ACCEPTED MANUSCRIPT then combined with the information on component sizing, the optimal total system pricing
234
could be computed.
235
Although the decision vector of stage two was 12 times as long as stage one (the whole year
236
instead of a typical month), and the iterative approach required multiple runs of stage two
237
for the local search, stage two optimization benefited largely from the fixed component
238
sizing, making bound setting and variable relationship generation significantly more efficient.
239
In addition, the fixed sizing component allowed better utilization of equality constraints,
240
compared to inequality constraints, they were overall much easier to optimize due to the
241
absence of slack variables. As a result, the solving of stage two problems greatly reduced the
242
overall memory requirement and significantly improved overall problem solving time.
SC
243
RI PT
233
3.2.1. Impact of loss of power supply probability (LPSP) on system sizing Nafeh [9] discussed on an additional aspect in his study, the loss of power supply probability.
245
This variable is set at the start of the optimization run, indicating the maximum allowed
246
intervals where the load was not met. Whereas Nafeh qualitatively discussed the potential
247
influence of LPSP, the actual impact was not quantified. This paper attempted to quantify
248
the impact of LPSP, along with other demand management strategies, through a
249
computationally light algorithm.
250
Ould Bilal et al. [17] modeled a PV-WT-BA system with similar constraints and the
251
minimization on system annualized cost, while maintaining a maximum LPSP level. Three
252
vastly different load profiles were applied to the model and the resulting annualized costs
253
were compared. It was shown that the profile with the least amount of fluctuation was best
254
handled by the hybrid system, requiring least amount of batteries and had the lowest
255
annualized cost.
256
To improve the performance and compensate the drawback of supply intermittency in
257
renewable energy systems, multiple layers of load with different supply priorities could be
258
implemented. Similar to load sensitive facilities such as a hospital, a critical load layer
259
represents the highest priority, where it should be fulfilled immediately without any delay. A
260
secondary layer of load represents demand of a less importance, where lower levels of LPSP
261
could be incorporated and allowed up to a certain degree. The third and last layer
262
represents the least critical demand type that allows load shifting or demand response. An
263
example of the lowest priority load is an electric storage hot water system, ideally the water
264
storage tank would consume electricity during midnight and have a full tank of water ready
AC C
EP
TE D
M AN U
244
10
ACCEPTED MANUSCRIPT at the start of the day, as long as a full tank of hot water is ready at the start of the next day,
266
the system has freedom to choose the optimized heating intervals that are most suitable
267
and economical. This concept would make the load model more flexible, giving the hybrid
268
system more leverage in meeting the demands, at the same time improve the overall energy
269
availability factor and overall system cost.
270
3.2.2. Model construction and solver software
RI PT
265
The model described above was constructed in MATLAB 2017a, attempted to be solved with
272
the MATLAB Global Optimization Toolbox, however the both solvers could not return
273
converged results in reasonable time, with each scenario run taking more than twelve hours
274
to converge when run on a PC with Intel Core i5-4570 CPU, 8GB RAM and 64-bit Windows 7
275
operation system. IBM ILOG CPLEX Optimization Studio 12.7.0 and Gurobi Optimizer 7.0
276
were introduced to the model in speeding up the solving. After many trials Gurobi optimizer
277
7.0 was select to be the optimization solver for the models discussed in this paper, due to its
278
superior speed and compatibility with MATLAB 2017a for this specific optimization problem.
279
A tolerance level of 0.5% was set to be a termination condition for both solvers, meaning
280
that the solving for problems would terminate if a feasible solution is found to be within 0.5%
281
difference in objective value to the theoretical optimum.
283 284
M AN U
3.2.3. Photovoltaic modeling
TE D
282
SC
271
Solar energy is the first renewable generation source used in the model. Traditionally, the
detailed modeling of PV systems uses short circuit current I and open circuit voltage V ,
like described by Zhou et al. [18]. To convert solar irradiance to electricity generation, a
286
photovoltaic model was needed, we used the California Energy Commission (CEC) 6
287
parameter Photovoltaic module model for its solar output calculation[19], with maximum
288
power point tracking system was considered to be operating at all times in the system in
289
order to provide the maximum electricity generation.
290
The electrical behavior of the equivalent circuit is described as
AC C
EP
285
= −
291 292 293
+ + − 1 −
Where is the output current, is the Light current, is the diode reverse saturation
current, is the series resistance, is the shunt (parallel) resistance, is the Boltzmann constant, is the absolute temperature of the PN junction, is the electron charge and
11
ACCEPTED MANUSCRIPT 294
is the number of cells in series. For detailed model specifications, we adopted the
295
specification of LG300N1-B3 panel module was selected for the model as it is one of the
296
highly regarded panel in the Australian market.
297
3.2.4. Wind turbine modeling Wind energy is the second renewable generation source in the model, it is included to give
299
the optimizer more freedom in selecting the appropriate system combination for the
300
electrical demand, also to supplement the lack of PV generation during night time. The
301
power output and electricity generation from wind turbines is a function of the blade shape,
302
pitch angle, blade swept radius and the rotational speed of the rotor unit [20].
M AN U
SC
RI PT
298
TE D
303
Figure 3. Typical power curve of wind energy conversion system [20]
305
For the model discussed in this paper, the specification parameters from Pitchwind 14m
306
30kW modeules were utilized, where the energy generation is mainly defined by four key
308
& − '& WT" = #$% & − '&
parameters, the cut-in wind speed, V( ,, the rated wind speed ,, the cut-out wind speed
+ and the rated power output P-.
AC C
307
EP
304
12
SC
RI PT
ACCEPTED MANUSCRIPT
309
Figure 4Pitchwind 14m 30kW generation curve
311
Similar to PV configurations, a MPPT process was also proposed for the wind turbine output
312
power generation [21]. The load dump included in the system schematic allowed generation
313
to exceed demand, therefore wind turbines were configured to generate maximum power at
314
all times. This practice avoided the additional operational control required on power
315
generation, reducing the complexity of the modeling problem and sped up optimal solution
316
finding.
TE D
317
M AN U
310
3.2.5. Diesel generator modeling
A diesel generator was included in the model to assist in peak demand times and periods
319
without renewable power generation. It was selected due to its portability and
320
independency on energy infrastructure. In addition, as a back-power application, diesel
321
generators are responsive to accept and supply electrical loads, making them good power
322
supply candidates in peak demand conditions. This component was therefore considered in
323
the model, with the purpose of preventing renewable energy system and battery
324
components from being oversized and underutilized due to peak demands requirements.
325
Diesel generators are more fuel efficient at higher loads, so in this model the diesel
326
generator is set to have a maximum generation efficiency of 32% at full load [22]. Similar to
327
most other fuel combustion engines, the part load engine efficiency are worse than rated. At
328
40% load the generation capacity decreased significantly to 27%. Although it is possible to
329
run diesel generators at even lower engine loads, the high fuel consumption makes it a poor
330
option in most scenarios. Moreover, due to the drastic change in efficiency between start-up
AC C
EP
318
13
ACCEPTED MANUSCRIPT and 40% load, piece wise linearization would require multiple segments to reflect the
332
efficiency characteristics, further slowing down the optimization solving process. Therefore
333
it was set in the model that diesel generators would not be utilized below 40% rated output.
RI PT
331
334
336
Figure 5. Typical 50kW diesel generator efficiency with output.
3.2.6. Inverter modeling
SC
335
Commercial inverters usually have a consistent efficiency throughout partial loads.
338
Nowadays, a Pulse Width Modulation (PWM) controller is often present in a power
339
converter to supply the AC load supply from a DC sources (DC to AC inverter) and also to
340
supply AC load from the DC bus (AC to DC rectifier) [23] The minor change in output
341
efficiency was insignificant towards the model, but also added unnecessary programming
342
complexity. Therefore in this paper the SMA Sunny Boy inverter was utilized, with a
343
weighted overall efficiency of 96%, and for conservative purposes, a fixed efficiency of 95%
344
was adopted in the model[24].
AC C
EP
TE D
M AN U
337
345 346 347
Figure 6. Power converter efficiency for the commercial SMA Sunny Boy inverter [24].
3.2.7. Rectifier modeling 14
ACCEPTED MANUSCRIPT In the rare cases of converting AC power to DC power (e.g. diesel generator charging the
349
battery storage system), an AC to DC rectifier would be required. Like the inverters,
350
commercial rectifiers have stable performance on output efficiency. The model in this paper
351
assumed the rectifier to perform similarly to the commercially available product GE
352
CAR3012TE rectifier [25], with a fixed efficiency of 93% when outputting 230 VAC.
353 354 355
M AN U
SC
RI PT
348
Figure 7GE CAR3012 TE Rectifier electrical specification
3.2.8. Battery modeling
Considering the nature of the problem, stand-alone systems are generally designed for rural
357
and areas outside electricity infrastructure, a well-studied energy storage device with
358
minimum maintenance should be selected. Compared to other battery technologies, Li-ion
359
batteries have high output voltage, and long cycle life [26]. These properties are highly
360
desirable for the stand-alone micro-grid, as high power output could assist in solving the
361
problems caused by renewable generation intermittency and long cycle life would be highly
362
beneficial during frequent charging and discharging events scheduled by the controller.
363
Being a well-developed technology Lithium-ion batteries were selected as the energy
364
storage device in the model, for its low maintenance and slow self-discharging. System
365
maintenance on the micro-grid system in rural areas is often expensive due to
366
inconvenience in logistics and transportation, low maintenance translated to a sizable saving
367
in operational expenditure. Similarly, in remote areas, batteries play an important role in
368
supporting renewable generation sources in fulfilling the local electricity demand. Therefore
369
battery cells that have fast and efficient charging/discharging cycles are more suited for the
370
task. Lithium-ion batteries are ideal for stand-alone systems that utilize renewable energy
371
sources [27] [28], because they could be fast discharged and charged without sacrificing
372
much efficiency, compensating the intermittent and volatile nature of renewable energy
AC C
EP
TE D
356
15
ACCEPTED MANUSCRIPT availability. The commercially available product BMZ ESS 7.0 Li-Ion NMC battery was used as
374
a reference.
SC
RI PT
373
375 Figure 8 BMZ ESS 7.0 Li-Ion system specification
377
However, some of the specifications of BMZ ESS 7.0 were not strictly enforced, such as the
378
max discharge current and max discharge power. The battery storage unit in the model was
379
allowed to rapidly charge and discharge, in order to assist the dire electricity condition seen
380
by the micro grid in extreme events. However, heavy efficiency penalty was applied to rapid
381
charge/discharge action to discourage such act as much as possible. The detailed
382
formulation is outlined in the piece wise linearization section later in this paper.
TE D
383
M AN U
376
3.2.9. Load modeling
The impact of LPSP on levelized system cost indicated an interesting issue from a difference
385
perspective, where the marginal cost associated in boosting the availability of renewable
386
energy systems with battery storage. Unlike fossil fuel based generators, renewable energy
387
systems have a lower availability index, even with the assistance of battery storage devices,
388
and full availability is often translated to system over-sizing under nominal operational
389
conditions. By implementing the LPSP index, the model was transformed into an effective
390
sensitivity testing tool that evaluated the marginal investment return on different levels of
391
LPSP. Yang’s model [5] provided a good insight for real world applications and economic
392
sensitive projects, that under certain scenarios it might be more desirable to compromise
393
full availability for lower overall system cost.
394
Trading electricity availability for reduced system cost is an attractive approach, but there
395
could be a potential incomprehensible aspect during loss of power supply periods. It was
396
likely that yang’s model compared the electricity supply and demand for each time-series
AC C
EP
384
16
ACCEPTED MANUSCRIPT data point, flagging loss of power supply if the demand was greater than the supply. With no
398
other control parameters on loss of power supply periods, it was likely that the GA solver
399
assigned zero electricity supply for those periods since from the optimizer’s perspective, any
400
supplied electricity supply during time points that failed to meet the demand could be better
401
used in other time points. Following the same logic, the intermittently generated electricity
402
from PV and WT would be directed towards the battery instead of serving the load. For
403
example, discussed by Yang et al. [5], a telecommunication tower was set as the energy
404
consumer, therefore it was reasonable for the GA optimizer to fully blackout the load under
405
allowed LPSP intervals. However, depending on the actual scenario this could potentially
406
generate problems, because full black out for every consumer could be extremely risky and
407
undesired (e.g. power failure to telecommunication and life supporting medical equipment).
408
In addition, it could be a wasteful act to charge the battery when immediate loads were
409
present, due to the extra losses incurred in the charging process (battery charging loss,
410
inverter loss if the battery charges from AC instead of DC, etc). Yang’s proposed method was
411
applied to a rural telecommunication tower project design [29], where a continuous 1.5 kW
412
electricity consumption was assumed as the demand. Although compromises were made
413
due to insufficient roof area for the PV panels, and the sub-optimal WT installation height
414
due to the prevention from potential natural disaster, the hybrid system operated well with
415
battery SoC above 50% at 90% of the time. The project data was monitored for one year and
416
over-discharge situations seldom occurred.
SC
M AN U
TE D
417
RI PT
397
3.3. Problem formulation
The model was constructed to include decision variables for both component sizing and
419
operational scheduling, see the Nomenclature section for abbreviations.
421 422 423
Let variable /0% denote the diesel generator generation in kWh, during time 1. When turned
AC C
420
EP
418
on, the output generation was limited to be at least 40% of the rated capacity, for better
efficiency and easier model construction. Therefore /0% was set as a semi-continuous variable that could take a value between a set range, or zero. /0% = 2
424 425 426
0.4dgs ≤ /0% ≤ dgs, :ℎ <:=1>ℎ / ? 0 , ?1ℎ @:=<
(1)
Let variable //%,A% denote the deferrable demand segmentation of deferrable electrical
demand BB% . //%,A% was set to be a special ordered set one for every 1, only allowing up to one variable to be non-zero.
17
ACCEPTED MANUSCRIPT EF
∀1, BB% = D //%,A% A%GH
∀1, ?IJ K 1? ? //%,A% >L M ?N @?
(2)
The unmet demand is formulated as the sum of instances where the electrical demand was
428
not fully fulfilled (LPSP hours). The maximum allowance of LPSP hours were set in the model
429
to reflect the minimum reliability of the microgrid.
Let variable =K/% denote the supply status of non-deferrable electrical demand /% .
431 432
1, 0,
:ℎ /% =< KOKIO=II / ?1ℎ @:=<
SC
=K/% = 2
allowance PB.
F
D =K/% ≤ PB %GH
433
TE D
≤ B% ∗ =K/% , = 0,
=K/" = 1 =K/" = 0
(5)
Let variable % denote the battery state of charge at time t, the battery state of charge
cannot exceed the boundaries of 0 and Q, where Q is the maximum battery capacity.
EP
435
(4)
In addition, variable K/% denotes the actual value of unfulfilled non-deferrable demand ∀1, K/% → 2
434
(3)
The total instances of unmet non-deferrable demand has to be smaller than the maximum
M AN U
430
RI PT
427
∀1, 0 ≤ % ≤ U
(6)
In addition, the battery was assumed to have 50% state of charging during initialization, this
437
would make the battery useful from the start of the model instead of only coming into effect
438
after being charged. However the energy during initialization was also requested at the end
439
of the time-series analysis, making sure no extra resource was given to the model.
440 441 442
AC C
436
1 qW ≤ U 2 1 q. ≤ U 2
(7) (8)
Let variable />% denote the change in state of charge at time 1. The change in state of
change was calculated as the difference in state of charge between the previous hour
(1 − 1) and present hour. Therefore, the change in state of charge is seen as an energy
18
ACCEPTED MANUSCRIPT 443
supply in the model, as positive figures reflect battery discharging and negative figures
444
reflect battery charging
1≠1 1=1
(9)
Let variable I% denote the electrical energy loss during battery charging and discharging ∀1, ql" = ^
/% ∗ (1 − _>`a ), /% ∗ (1 − _/`a ),
ML11 @J >ℎL@0=0 battery discharging
RI PT
445
− % , ∀1, /% = 2 %[H 0,
(10)
446
Total energy supply at all time-series point must greater than or equal to the total demand
447
of the same hour EF
SC
∀1, /0% + _k' ∗ (/% + #% ) + l% + D //%,A% A%GH
448
3.4. Piece-wise linearization
(11)
M AN U
≥ B% + BB% + I% − K/% ∗ =K/%
Two of the main components in the model, namely the diesel generator and the lithium-ion
450
battery storage system have non-linear characteristics that have to be addressed. Due to the
451
limitations of a linear model, piece-wise linearization of non-linear functions and
452
relationships were performed to enable them being accepted by the linear solver.
453
TE D
449
3.4.1. Diesel generator
The diesel generator operating efficiency curve was cut into three segments, each segment
455
had a unique efficiency, the overall fuel consumption used in the objective function was
456
calculated as the sum of fuel generation all segments, which was formulated as a special
457
ordered set 1, where constraints were set to only allow up to one segment to be non-zero,
458
in order to avoid the conceptual error of diesel generator operating on multiple output
459
efficiencies
AC C
EP
454
_`n
27%, = o30%, 32%,
∀1, O% =
/0% _`n
0.4/0< ≤ /0 < 0.6/0< 0.6/0< ≤ /0 < 0.8/0< 0.8/0< ≤ /0 ≤ /0<
(12)
(13)
460
The efficiency of diesel generator was cut into three segments, namely between 40% and 60%
461
load, between 60% and 80% and above 80% load. Segment under 40% load was avoided 19
ACCEPTED MANUSCRIPT from the system design, due to the inefficient operation. This was achieved by assigning a
463
heavy penalty for operational segment below 40%, achieved by an unrealistic low output
464
efficiency. This act made low load generation strictly worse than higher load (generating
465
little electricity while consuming equal or greater amount of diesel), hence eliminating the
466
potential of low load generation scenarios.
RI PT
462
SC
467 Figure 9 Diesel generator efficiency linearization
469
The model formulation of the piece-wise linearization on the diesel generator involved
470
auxiliary parameters and utilized the concept of big-M method.
471
Two constraints govern the behavior of each of the three linear segments for diesel
472
M AN U
468
generator, on every time-series point 1. Let /0%, denote the electricity generation from the
474
diesel generator segment < at time 1, Boolean variable =vw%, denote the operating status of
475
be a positive value that is much larger than the upper-bound of all the variables in the model.
the diesel generator segment < at time <, where 1 represents on and 0 represents off, and x
TE D
473
/0%,H ≥ 0.4/0< − z1 − =vw%,H { ∗ x
y/0%,& ≥ 0.6/0< − z1 − =vw%,& { ∗ x /0%,| ≥ 0.8/0< − z1 − =vw%,| { ∗ x
477
478 479
480 481
(15)
/0%,H ≥ 0.4/0< − x
(16)
EP
∀<, /0%, ≤ =vw%, ∗ x
Variable set =vw limits the value of the segment output, taking the first segment as an
example, when =vw%,H = 0, the constraints above become:
AC C
476
(14)
/0%,H ≤ 0
(17)
Due to the large number nature of variable x, 0.4/0< − x will become a negative number and these constraints limit the value of /0%,H to be:
0.4/0< − x ≤ /0%,H ≤ 0
(18)
Since the value range of generator output is set to be non-negative in the bound setting, the
only feasible value for /0%,H would be 0. On the other hand, when =vw%,H = 1,
20
ACCEPTED MANUSCRIPT /0%,H ≥ 0.4/0<
(19)
/0%,H ≤ x
(20)
483
The segment output /0%,H has the freedom to select any value between 40% of the rated
484
generator output and large number x. Since the range setting also forbids the segment generation output to be greater than the rated diesel generator output, therefore the
485
effective variable range is:
488 489
RI PT
487
(21)
Under the same principle, the variable range for the second and third segments, when their
corresponding =vw variable is 1, are:
0.6/0< ≤ /0%,& ≤ /0< 0.8/0< ≤ /0%,| ≤ /0<
SC
486
0.4/0< ≤ /0%,H ≤ /0<
(22) (23)
Few extra constraints were needed to connect the three segmented pieces together. Firstly the special ordered set 1 constraints on =vw variables prevents generation double counting:
M AN U
482
∀1, =vw%,H + =vw%,& + =vw%,| ≤ 1
Next, the total
diesel fuel consumption for time t were expressed
AC C
as:
/0% = /0%,H + /0%,& + /0%,|
EP
and total
TE D
generator
O>% =
/0%,H /0%,& /0%,| + + _`n,H _`n,& _`n,|
(24)
(25)
(26)
490
It is worth noting that, in constraints 27 to 28, the upper limit of each segment was not
491
enforced onto the output variables. This was done intetionally to reduce the number of
492
constraints in the model, although eneration output variables in segment one could
493
technically have a value greater than 60% of the rated capacity, but this would be a strictly
494
worse solution compared to using the second segment to generate equal amount, due to
495
the improved fuel efficiency in the later segments. Since the objective of the model is to
496
minimize overall system cost, the optimization solver will always prioritize segments with
21
ACCEPTED MANUSCRIPT 497
higher fuel efficiency, and only use segments with lower efficiency when small amount of
498
electricity is required.
499
3.4.2. Battery storage system The construction of the battery storage system used the same principle as the diesel
501
generator constraint setting, with the exception of the batteries can have bi-directional flow
502
with different efficiencies. Therefore battery charging and discharging efficiency were split
503
into two parts and done separately. Four segments were created to represent charging
504
efficiencies and another four segments for discharging. Therefore a total of 8 segments were
505
applied for linearizing the battery performance. The model was constructed using hourly
modeling time-series resolution, therefore the c-rate of batteries could not go beyond } at
SC
506
RI PT
500
any interval.
508
For formulation purposes, we only considered the useful battery capacity to simplify the
509
model construction, however the full component cost was included and normalized to the
510
effective battery capacity for cost calculations. The battery module in this model used the
511
popular Panasonic NCR18650 series for the performance benchmark. The battery would
512
perform better and last longer with longer charging and discharging time, we transformed
513
these characteristics into charging and discharging efficiency to encourage slow
514
charge/discharge [30], due to the fact the one year time-span in this model was not enough
515
to reflect the long term capacity degradation. The discharging efficiency of the battery could
516
be represented as
TE D
M AN U
507
517 518 519
AC C
EP
85%, 87%,
_/`a =
90%, ~93%,
}` < / ≤ }` 1.5 }` }` < / ≤ 2 1.5 }` }` < / ≤ 2.5 2 }` 0 ≤ / ≤ 2.5
(29)
The negative of / was taken during charging loss calculations, this was due to the charging
aspect of the battery would have / take negative values
22
ACCEPTED MANUSCRIPT
520 521
90%, ~93%,
*charts of real vs linearized segments
} < −/ ≤ } 1.5 } } < −/ ≤ 2 1.5 } } < −/ ≤ 2.5 2 } 0 ≤ −/ ≤ 2.5
3.5. Residential electrical demand
(30)
RI PT
_>`a =
85%, 87%,
Isolated islands do not have the sophistication of detailed infrastructure control and
523
monitoring as large scale electricity network. The hourly electricity demand profile was
524
difficult to obtain. Considering the difficulties, the residential electrical demand profile used
525
In the model referenced, the average electricity demand of an urban residential precinct,
526
mainly in the Greater Sydney region in New South Wales, Australia.
527
To determine the deferrable electricity demand, the household electrical demand was
528
disaggregated down to individual appliances, where their electricity consumptions and
529
usage profiles were mapped out. In the end, several appliances were given the flexibility of
530
load deferring, namely the washing machine, cloth dryer, dish washer and the hot water
531
system. Other appliances that return an immediate amenity, such as lighting and air-
532
conditioning systems were not allowed from load deferring and have their demands delayed,
533
in order to maintain the resident comfort level.
534
The load profile of simulated household appliance usage on a 24 hour period was generated
535
using the methodologies listed by Richardson et al. [31] and benchmarked using the
536
commercial software CCAP Precinct [32]. The nature of stand-alone micro-grids being
537
disconnected to the well monitored electricity infrastructure, it was difficult for us to obtain
538
all the necessary parameters for simulating the electricity demand generated on the islands.
539
As an alternative, synthetic energy use behaviors of coastal Sydney areas were instead
540
applied into the modeling of demand profiles the Eastern Australian islands.
541
AC C
EP
TE D
M AN U
SC
522
3.6. Economic criteria and cost accounting
542
The modeling of MILP on minimizing the overall system cost requires each system
543
component to have levelized costs. In this paper, cost parameter comprises capital cost,
544
maintenance cost, component lifespan and operational fuel cost [33]. Since the analysis
545
performed in this paper did not extend to the full lifetime of each system component, it was
23
ACCEPTED MANUSCRIPT 546
important to normalize the component cost to a leveled platform. As a result, each
547
component cost was reduced down to an annual figure, effectively the cost of an annual
costs, taking into account the Consumer Price Index of Australia in recent years. P
= #'k%
550
@(1 + @) + #A%' (1 + @) − 1
(31)
RI PT
549
repayment of the component. The annual i rate @ of 5% [34] was considered for all system
The unit price of each of the components modeled in this study is shown in Table 1 Technology
Cost ($/kW, $/kWh or $/L) 1,800 3,800 800 1,333 1.25
(Years)
20 20 20 (assuming minimal usage) 5 -
M AN U
Photovoltaic Panel Wind Turbine Diesel Generator Battery Diesel (fuel)
Lifetime
SC
548
Table 1: Technology component pricing
552
4. Results and Discussion
553
Programmable household appliance loads (e.g. hot water system, washing machine, dish
554
washer and water pumping) with the ability to have the operation delayed with minimal
555
interference to the everyday life of the residents were extracted from the total electricity
556
demand and modeled as a second demand layer. A deferrable time allowance was applied
557
to the load and the optimal system configuration was then optimized on minimal levelized
558
electricity cost.
559
To explore the potential of load deferring, two scenarios were firstly modeled and compared;
560
one without any freedom of load deferring and the other with up to 10 hours of load
561
deferring allowance. Under normal residential use, appliance loads reach their peak during
562
night-time, between dusk and dawn of the next day. In addition, commercially available
563
electric hot water systems are also programmed to operate overnight, where a large amount
564
of electricity would be consumed. In contrast, to PV generation characteristics where all the
565
generations happen during the day, the big misalignment between appliances and
566
photovoltaic often make the photovoltaic generation utilization much more difficult. By
567
allowing the load to be deferred by up to 10 hours in the load deferring scenario, a
568
significant portion of the deferrable load was delayed and fulfilled the next morning, when
569
PV generation was directly available. This act of load deferring minimized the reliance on
AC C
EP
TE D
551
24
ACCEPTED MANUSCRIPT battery storage systems by shifting blocks of electrical demand towards generation, it
571
created better alignment between the consumption and generation profile, hence
572
enhancing the effectiveness and utilization of renewable generation sources with little
573
hardware investments needed, yielding a noticeable reduction in overall system cost.
M AN U
SC
Electricity demand - kW
RI PT
570
574
Figure 10 Original and rescheduled average electricity demand profile
576
Observing the electricity demand profile result in Figure 10, load deferring was seen to have
577
a major impact in rescheduling the electrical demand for other system components. We
578
computed and compared the original daily average demand profile against the rescheduled
579
demand profile in the load deferring controller
580
The original demand profile had high night time demand, mainly due to the consumer
581
electronics consumption outside working hours and the overnight water heating from an
582
electric hot water system.
EP
TE D
575
Electricity demand - kW
AC C
583
584 585
Figure 11 Original deferrable demand profile vs. Rescheduled deferrable demand profile
25
ACCEPTED MANUSCRIPT Shown in Figure 11, 10 hours of load deferring was allowed for this scenario run, the night
587
time demand was heavily shifted until the next morning, to better align with the expected
588
electricity generation from PV. Furthermore, a sizable average peak demand reduction of 30%
589
was also observed in the results, which indicated the successful attempt of load
590
management in the form of demand smoothing, as well as the potential in hardware sizing
591
reduction and improved network resilience.
592
Next, the deferrable load, representing approximately 48% of the overall electricity demand
593
were extracted and studied. Comparing the two figures, deferrable demands were shifted
594
and redistributed towards the middle of the day to best utilize the PV generation. In addition,
595
the deferred electrical demand resembled the PV generation profile during periods with
596
solar irradiation, indicating the deferrable load controller and system optimizer attempting
597
to restructure demands to best follow the PV generation profile, in order to reduce reliance
598
on battery storage systems and diesel generators.
SC
M AN U
599
RI PT
586
4.1. New establishment / Component replacement
After validating the significance of load deferring on electricity demand profiles, the model
601
was run on several scenarios, attempting to investigate the effectiveness of load deferring,
602
when coupled with programmed LPSP, one of the most commonly adopted demand
603
management strategies. Sizing composition of the optimal system configuration under
604
different parameter settings was graphed. Diesel generator as an electricity generation
605
source was sized smaller and not favored when the demand profiles got relaxed and
606
smoothed. The strength of diesel generators lies in dealing with spiky and intermittent loads,
607
where it also has its obvious drawback of fossil fuel reliance, hence its high operating cost.
608
The steady decrease of diesel generator component utilization was observed as both
609
demand management strategies strengthened and the electricity demand profile were
610
increasingly smoothed and rescheduled to better fit the renewable energy generation
611
profile. In other words, demand management strategies covered the strength of diesel
612
generators, where its drawbacks were left unchanged. As a result, as load deferring and LPSP
613
gained more control in demand rescheduling, diesel generator became less and less
614
attractive as a electricity generator component, hence getting its sized reduced to only
615
provide assistance in adverse situations.
AC C
EP
TE D
600
26
SC
RI PT
Diesel/PV/Wind - kW Battery - kWh
ACCEPTED MANUSCRIPT
M AN U
616
Figure 12 Component sizing comparison on new precinct establishment
618
It was also shown in the results that whereas wind was the dominant component in the
619
system, its size was relatively unchanged in scenarios with higher loss of electricity and
620
deferrable load. Although on a unit pricing basis, wind is more than twice the cost than solar
621
PV, but its less fluctuating generation profile still benefited more to a shifted and relaxed
622
residential load, it could be seen from the graph that wind turbines were heavily sized in all
623
optimal system configuration setups.
AC C
EP
TE D
617
27
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
624
Figure 13.System cost reduction against scenario without any LPSP and load deferring.
AC C
EP
TE D
625
626 627
Figure 14 Levelized electricity cost reduction comparison for different scenarios
28
ACCEPTED MANUSCRIPT The levelized system cost reduction was then calculated. Demand management strategies
629
achieved an overall reduction up to 27%, when 5% LPSP was scheduled and up to 10 hours
630
of load deferring were allowed. From the comparison in Figure 14, the demand side
631
strategies were more effective in their initial take ups, and suffered diminishing returns
632
when both strategies were granted more control over the demand profile. Therefore it was
633
clearly concluded that the impact of demand management strategies would quickly
634
converge, where additional freedom and power granted to the demand management
635
strategies might not be a good investment, especially taking into account resident amenity.
636
However, demand management strategies proved to be extremely effective at low values,
637
when the original electricity demand profile provided the most opportunities for demand
638
rescheduling.
639
The surface plot in Figure 14 presented a clear view of the system cost reduction at different
640
levels of LPSP and load deferring. Such surface plots have been presented and discussed in
641
previous literatures to illustrate system performances on multi parameter configurations
642
[35].
643
Reading from the graph, LPSP shown a stronger total system cost reduction, without the
644
assistance of load deferring, 5% of LPSP could reduce the overall system cost by 16%.
645
However, viewing from the residents’ perspective, high LPSP is very undesirable and it will
646
severely affect residents’ amenity and productivity. At 5% LPSP, a typical resident could
647
expect up to 438 hours of electricity blackout each year, heavily impacting their living quality
648
and productivity. Therefore, a less disruptive alternate solution to reduce levelized energy
649
cost would be much preferred by the consumers.
650
The other key parameter in the analysis was load deferring. At 10 hours allowance alone,
651
load deferring provided 11% system cost reduction against the reference scenario. While
652
being less effective than LPSP, load deferring was a much less disruptive load management
653
strategy than LPSP, as it only involved the non-critical residential electricity loads being
654
potentially delayed, examples being pressing the start button on the washing machine and
655
have the laundries done after 5 hours, or having the daily water heating schedule moved
656
away from midnight hours.
657
The overall surface had a gradual reduction in slope towards high values of both LPSP and
658
load deferring allowance, however the excess interference in resident amenity could create
659
bigger social problems. Thus we suggest the most effective approach could be found in a
660
combination of LPSP and deferrable load, both capped at relatively low values, this
AC C
EP
TE D
M AN U
SC
RI PT
628
29
ACCEPTED MANUSCRIPT configuration would provide a relaxed generation and battery operation environment for
662
the stand-alone hybrid systems, hence reducing the total system cost without too much
663
performance loss and resident discomfort.
664
The optimal solutions obtained by MILP to minimize the levelized electricity cost in a normal
665
operation year. Each point of the graph is associated with a set of input decision variables
666
including PV sizing, wind turbine sizing, diesel generator sizing, battery sizing and optimal
667
battery management. It is shown that allowing loads to be deferred had a significant impact
668
on the system, without any compromise on the total demand. However the savings from
669
deferrable load shown a diminishing return, allowing loads to be deferred for long time (10+
670
hours) did not show significant improvement over a small deferrable hour (e.g. 2 hrs).
671
Observing the plotted results, there were non-smooth points on the surface, typically along
672
the lines of 2.5% LPSP allowance. Multiple runs of the model were performed with different
673
parameters, such as adjusting the frequency of heuristic guesses and utilizing different linear
674
programming methods (simplex, interior point etc.) to yield the same result. This could be
675
the result of the model solution trapped in a local optimum or the absence of feasible
676
solutions close to the set tolerance limit.
SC
M AN U
4.2. Network upgrade of existing system
TE D
677
RI PT
661
This section simulated network upgrade of an existing stand-alone micro-grid, whereas
679
diesel generator was the sole electricity supplier. Hence the diesel generator was sized big
680
enough to supply the electrical demand without any other assistance. A network upgrade
681
was then proposed to reduce the system cost of the micro grid by introducing renewable
682
energy sources and energy storage systems.
683
The diesel generator capable of supplying all electricity demands was set to be present in
684
every scenario, with no capital cost associated. Similar optimization runs were performed to
685
explore the system cost saving opportunities in load deferring and allowed loss of electricity.
AC C
EP
678
30
686 Figure 15 Component sizing comparison for network upgrade scenario
688
AC C
EP
TE D
M AN U
687
SC
RI PT
Diesel/PV/Wind - kW Battery - kWh
ACCEPTED MANUSCRIPT
689
Figure 16 Levelized electricity cost reduction on various scenarios on network upgrade
690
The difference in the levelized electricity cost reduction across all modeled scenarios in the
691
network upgrade scenario were observed to be much smaller than the new development
692
establishment scenario. Given the presence of the diesel generator, demand side
693
management strategies were shown to be less influencing in capital cost saving from system
31
ACCEPTED MANUSCRIPT 694
component sizing reductions. The existing diesel plant proved to be a reliable backbone for
695
the electricity system, where adding new RES components became less cost effective
696
compared to the new establishment scenario modeled in the previous section. The
697
additional operational cost saving mainly came from the reduction of fuel usage, which was
698
only a modest piece in influencing the overall levelized electricity price. 4.3. Compare to previous studies and relevant references
RI PT
699
Compared to studies in other literatures that constructed non-linear models and solved with
701
evolution based algorithms, the approach described in this paper achieved similar results
702
with linearized constraints. Non-linear models described in other studies [36] [37] utilize a
703
bottom-up approach where the optimization model would only cover a short period of time
704
(hourly resolution for 24 hours to a week), and then the results are post-analysed to reflect
705
the likely annual behavior. This paper took the approach of parameter linearization,
706
sacrificing non-linear relationship accuracy for better computational speed to afford longer
707
time interval simulations, which could be the preferable approach to support decision
708
making for areas located in high latitude zones, such as Sydney, Australia discussed in this
709
paper. The annual coverage in time-series points reduces the error in finding the global
710
optimum configuration that rely on post-anlaysing short time interval results into annual
711
levelized cost. For most non-linear models, the post-model-analyse on week-long
712
optimization results to predict the optimum for the annual optimum could be relatively
713
straight forward in areas with a stable climate, such as countries near the equator, however
714
the same analyse becomes difficult and challenging, due to the high seasonal variability in
715
daylight hours and temperature in other areas around the globe.
M AN U
TE D
EP
4.4. Future research
AC C
716
SC
700
717
Component degradation was not modeled in details for the system and could be
718
investigated further. (e.g. diesel generator and battery were utilized less as load deferring
719
and loss of electricity increased, this could potentially translate to a reduced maintenance
720
cost and longer life time)
721
Impact of component pricing fluctuation on the system configuration, were not factored in
722
the analysis, however it would be another determining factor in component sizing, economic
723
sensitivity analysis could be undertaken to describe the likelihood of sizing differences when
724
capital costs change.
32
ACCEPTED MANUSCRIPT The impact of greenhouse gas emission was not discussed in the model, with the
726
introduction of RES, the fossil fuel usage in electricity generation would drop by a significant
727
amount, which would directly lower the greenhouse gas emission. As carbon emission
728
getting more attention globally, along with the increase in fuel price, most countries have
729
been considering a greener future development strategy, where the excess emission of
730
greenhouse gas would be penalized, where the use of RES gets incentivized. This shift in
731
government policies, although not quantitatively implemented in this model, would give RES
732
an even bigger advantage in economic terms, than traditional fossil fuel burning systems for
733
electricity generation.
RI PT
725
AC C
EP
TE D
M AN U
SC
734
33
ACCEPTED MANUSCRIPT
5. Conclusions
736
A two stage mixed-integer linear programming model was constructed and used to explore
737
the possibilities in optimizing the reduction of stand-alone renewable system costs with
738
diesel and energy storage. The results obtained reveal that load deferring is a cost-effective
739
and non-disruptive method of demand side control strategy for managing and adapting
740
residential electricity demand profiles to better align with renewable generation profiles. In
741
addition, the introduction of load deferring greatly reduced the battery capacity required in
742
the system, which was the biggest contributor towards lowering annual system costs.
743
Sizing composition charts were also developed to show the usefulness of wind turbines in
744
this particular stand-alone system, which allow load deferring and loss of power supply
745
probability. For the island scenario proposed in this paper, the nature of high latitude
746
created big seasonal variance in PV generation, coupled with the high wind speeds in coastal
747
regions, wind turbines have better capacity factors and availability than PV for the context of
748
this study, hence allowing them to be less reliant on batteries. This reliance reduction is
749
sufficient to overcome the drawback of high capital prices for wind turbines. It is expected
750
that this work will complement concurrent efforts in the field to improve the robustness and
751
computational cost-effectiveness of scheduling and sizing investigations.
SC
M AN U
TE D EP AC C
752
RI PT
735
34
ACCEPTED MANUSCRIPT
Nomenclature
l
,
l
l
lℎ
$ $/ l/&
¡¢,
Boolean
¥
Boolean
¦§ ¨
$
¨ª
ℎ@
«, «¬ «
% % %
«
%
«
%
®®¥¯
$
®°®®°
TE D
EP
¬ ¯
-
photovoltaic panel area idealized charging/discharging rate to fully charge/discharge the battery in one hour (energy losses considered separately in I`a term) electricity to supply deferrable electricity demand BB% , 1 hours after time 1 diesel generator capacity (rated) variation in battery state of charge between time 1 − 1 and 1, special case />W = 0 Total diesel generator fuel cost at t normalized diesel generator fuel cost diesel generator fuel consumption at time 1 global horizontal solar irradiation at time 1 boolean variable denoting the operation status of the diesel generator boolean variable denoting loss of electricity status at time 1 lifespan (in years) of the system component Auxiliary variable for piece-wise linearization constraints maximum time before any deferrable load must be fulfilled diesel generator efficiency for output segment < power converter (inverter/rectifier) efficiency photovoltaic panel system efficiency battery charging efficiency as a function of battery charge rate / battery discharging efficiency as a function of battery charge rate / annual component cost (including capital and maintenance costs, excluding fuel cost) modular capital cost of the system component annual modular maintenance cost of the system component electricity generation from photovoltaic panels at time 1 photovoltaic panel size maximum battery storage capacity battery state of charge at time 1 battery electrical loss during charging/discharging as a function of charging/discharging rate battery energy loss due to charging/discharging at time 1 annual financing interest and consumer price inflation total modeling interval unmet demand at time 1 due to allowed loss of electricity maximum allowed unmet demand over the modeling period. electricity generation from wind turbines at time 1
RI PT
m&
Meaning
SC
Units
M AN U
Symbol
$ $
AC C
753
¬± ²
l l l [0, 1]
¯
lℎ
¯ µ ª ¥
lℎ J L@ ℎ@ l
¶·
l
¸ª
l
35
wind turbine size
EP
TE D
M AN U
SC
RI PT
l
AC C
754
¹
ACCEPTED MANUSCRIPT
36
ACCEPTED MANUSCRIPT 755
References
756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802
1.
2.
RI PT
3.
De Groot, M., J. Forbes, and D. Nikolic. Demand response in isolated power systems. in 2013 Australasian Universities Power Engineering Conference, AUPEC 2013. 2013. Hobart, TAS. Fezai, S. and J. Belhadj. Load profile impact on a Stand-Alone Photovoltaic system. in 2016 7th International Renewable Energy Congress (IREC). 2016. Bilal, B.O., et al., Study of the Influence of Load Profile Variation on the Optimal Sizing of a Standalone Hybrid PV/Wind/Battery/Diesel System. Energy Procedia, 2013. 36: p. 1265-1275. Caisheng, W. and M.H. Nehrir, Power Management of a Stand-Alone Wind/Photovoltaic/Fuel Cell Energy System. Energy Conversion, IEEE Transactions on, 2008. 23(3): p. 957-967. Yang, H., et al., Optimal sizing method for stand-alone hybrid solar–wind system with LPSP technology by using genetic algorithm. Solar Energy, 2008. 82(4): p. 354-367. Merei, G., C. Berger, and D.U. Sauer, Optimization of an off-grid hybrid PV–Wind– Diesel system with different battery technologies using genetic algorithm. Solar Energy, 2013. 97: p. 460-473. Shang, C., D. Srinivasan, and T. Reindl. Joint generation and multiple demand scheduling in off-grid buildings. in IEEE International Conference on Building Energy Efficiency and Sustainable Technologies, ICBEST 2015. 2015. Institute of Electrical and Electronics Engineers Inc. Koutroulis, E., et al., Methodology for optimal sizing of stand-alone photovoltaic/wind-generator systems using genetic algorithms. Solar Energy, 2006. 80(9): p. 1072-1088. Nafeh, A.E.-S.A., Optimal Economical Sizing Of A PV-Wind Hybrid Energy System Using Genetic Algorithm. International Journal of Green Energy, 2011. 8(1): p. 25-43. Mazzola, S., M. Astolfi, and E. Macchi, A detailed model for the optimal management of a multigood microgrid. Applied Energy, 2015. 154: p. 862-873. Chen, S.X., H.B. Gooi, and M.Q. Wang, Sizing of Energy Storage for Microgrids. IEEE Transactions on Smart Grid, 2012. 3(1): p. 142-151. Morais, H., et al., Optimal scheduling of a renewable micro-grid in an isolated load area using mixed-integer linear programming. Renewable Energy, 2010. 35(1): p. 151-156. Dai, R. and M. Mesbahi, Optimal power generation and load management for offgrid hybrid power systems with renewable sources via mixed-integer programming. Energy Conversion and Management, 2013. 73: p. 234-244. Faxas-Guzmán, J., et al., Priority load control algorithm for optimal energy management in stand-alone photovoltaic systems. Renewable Energy, 2014. 68: p. 156-162. Bekele, G. and B. Palm, Feasibility study for a standalone solar–wind-based hybrid energy system for application in Ethiopia. Applied Energy, 2010. 87(2): p. 487-495. Bekele, G. and G. Tadesse, Feasibility study of small Hydro/PV/Wind hybrid system for off-grid rural electrification in Ethiopia. Applied Energy, 2011. 97: p. 5. Ould Bilal, B., et al., Optimal design of a hybrid solar–wind-battery system using the minimization of the annualized cost system and the minimization of the loss of power supply probability (LPSP). Renewable Energy, 2010. 35(10): p. 2388-2390. Zhou, W., H. Yang, and Z. Fang, A novel model for photovoltaic array performance prediction. Applied Energy, 2007. 84(12): p. 1187-1198.
4.
5.
SC
6.
9. 10. 11. 12.
AC C
13.
EP
8.
TE D
M AN U
7.
14.
15. 16. 17.
18.
37
ACCEPTED MANUSCRIPT
24.
25. 26. 27.
28. 29. 30.
31. 32.
33.
RI PT
23.
SC
22.
M AN U
21.
TE D
20.
Dobos, A.P. Improved Coefficient Calculator for the California Energy Commission 6 Parameter Photovoltaic Module Model. United States: USDOE Office of Energy Efficiency and Renewable Energy Solar Energy Technologies Program. Pallabazzer, R., Evaluation of wind-generator potentiality. Solar Energy, 1995. 55(1): p. 49-59. Koutroulis, E. and K. Kalaitzakis, Design of a maximum power tracking system for wind-energy-conversion applications. IEEE Transactions on Industrial Electronics, 2006. 53(2): p. 486-494. Silva, S.B., M.A.G. de Oliveira, and M.M. Severino, Economic evaluation and optimization of a photovoltaic–fuel cell–batteries hybrid system for use in the Brazilian Amazon. Energy Policy, 2010. 38(11): p. 6713-6723. Dufo-López, R., I.R. Cristóbal-Monreal, and J.M. Yusta, Stochastic-heuristic methodology for the optimisation of components and control variables of PV-winddiesel-battery stand-alone systems. Renewable Energy, 2016. 99: p. 919-935. AG, S.S.T. Sunny Boy Inverters. 2017; Available from: http://www.smaaustralia.com.au/products/solarinverters/sunny-boy-30-36-40-50.html#Overview244074. solutions, G.I. CAR Series of front-end rectifiers. 2017; Available from: http://www.geindustrial.com/products/embedded-power/carmpr. Dunn, B., H. Kamath, and J.-M. Tarascon, Electrical Energy Storage for the Grid: A Battery of Choices. Science, 2011. 334(6058): p. 928-935. Liu, X., et al., A new state-of-charge estimation method for electric vehicle lithiumion batteries based on multiple input parameter fitting model. International Journal of Energy Research, 2017: p. n/a-n/a. Huria, T., et al. High fidelity electrical model with thermal dependence for characterization and simulation of high power lithium battery cells. IEEE. Yang, H., Z. Wei, and L. Chengzhi, Optimal design and techno-economic analysis of a hybrid solar–wind power generation system. Applied Energy, 2009. 86(2): p. 163-169. Panasonic. Panasonic Lithium-Ion NCR18650. 2017 [cited 2017; Battery specification sheet]. Available from: https://na.industrial.panasonic.com/sites/default/pidsa/files/ncr18650b.pdf. Richardson, I., et al., Domestic electricity use: A high-resolution energy demand model. Energy and Buildings, 2010. 42(10): p. 1878-1887. Kinesis, CCAP Precinct. 2016. p. A strategic urban design software used to predict the environmental, economic and social impacts of residential, commercial and mixeduse developments. Rajanna, S. and R.P. Saini, Modeling of integrated renewable energy system for electrification of a remote area in India. Renewable Energy, 2016. 90: p. 175-187. Kalisch, D.W., Consumer Price Index, A.B.o. Statistics, Editor. 2017. Koutroulis, E. and D. Kolokotsa, Design optimization of desalination systems powersupplied by PV and W/G energy sources. Desalination, 2010. 258(1–3): p. 171-181. Kapfudza, M., N. Moorosi, and M. Thinyane. GA optimization of Hybrid Energy Systems for telecommunications in marginalised rural areas. in 2014 6th IEEE International Conference on Adaptive Science and Technology, ICAST 2014. 2015. IEEE Computer Society. Tutkun, N. and O. Can. Optimal load management in a low power off-grid windphotovoltaic microhybrid system. in 16th International Conference on Environment and Electrical Engineering, EEEIC 2016. 2016. Institute of Electrical and Electronics Engineers Inc.
EP
19.
AC C
803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851
34. 35. 36.
37.
852
38