Professional Documents
Culture Documents
Energy
journal homepage: www.elsevier.com/locate/energy
a r t i c l e i n f o a b s t r a c t
Article history: This paper proposes a new method to optimize the scheduling of off-grid systems composed of solar
Received 13 October 2020 panels, batteries, and diesel generators in real time. The approach takes into account the load and
Received in revised form irradiance forecasted values for the near future to determine the optimal power generation and the
13 January 2021
operation of the different energy sources which achieve the lowest cost while fulfilling the provided
Accepted 15 January 2021
Available online 23 January 2021
constraints. A real-time simulator is employed to run the simulations with a high degree of accuracy to
further validate the obtained results as well as to analyze the grid quality parameters (frequency, voltage,
and harmonics). To validate the effectiveness of the proposed forecast-based approach, the performance
Keywords:
Forecasting
from two benchmark algorithms commonly applied in these systems are also estimated. The final results
PV-Battery-diesel hybrid systems reveal that the proposed algorithm is able to achieve 5% cost savings with respect to the benchmark
Real-time simulator approaches while still fulfilling the grid quality constraints. The proposed method can then be applied for
Scheduling optimization real off-grid systems to further enhance their performance.
© 2021 Elsevier Ltd. All rights reserved.
Credit author statement from this manuscript was discussed. Thomas Reindl: The concep-
tualization of the work from this manuscript was discussed, The
Carlos D. Rodríguez-Gallegos: The conceptualization of the work funding acquisition belongs to, The supervision of this work was
from this manuscript was discussed, The data analysis was per- carried out. S K. Panda: The conceptualization of the work from this
formed, The investigation and methodology was performed, The manuscript was discussed, The funding acquisition belongs to, The
original draft was written by CDRG and reviewed and edited by all supervision of this work was carried out.
other co-authors. Lokesh Vinayagam: The conceptualization of the
work from this manuscript was discussed, The data analysis was 1. Introduction
performed, The investigation and methodology was performed.
Oktoviano Gandhi: The conceptualization of the work from this Dispatch strategy, also known as scheduling, refers to the
manuscript was discussed, The data analysis was performed, The operation and power control of the different energy sources in a
investigation and methodology was performed. Gokhan Mert Yagli: system. In this study, the sources correspond to the PV system,
The data analysis was performed, The investigation and method- battery system and diesel generators. It is then desired to control
ology was performed. Manuel S. Alvarez-Alvarado: The data anal- these sources to satisfy the defined constraints, such as the fulfill-
ysis was performed, The investigation and methodology was ment of the load demand, and to reduce the overall costs and CO2
performed. Dipti Srinivasan: The conceptualization of the work emissions, among other objectives.
With respect to PV hybrid systems, the literature presents
different works to optimize their dispatch strategy, typically to
* Corresponding author. Solar Energy Research Institute of Singapore (SERIS),
achieve cost reductions, e.g. Refs. [1e5] where different optimiza-
National University of Singapore (NUS), 117574, Singapore. tion algorithms are proposed (some of these presented in Ref. [6]).
E-mail address: carlos.rodriguez@nus.edu.sg (C.D. Rodríguez-Gallegos). From the previous references, as well as from many more that can
https://doi.org/10.1016/j.energy.2021.119918
0360-5442/© 2021 Elsevier Ltd. All rights reserved.
C.D. Rodríguez-Gallegos, L. Vinayagam, O. Gandhi et al. Energy 222 (2021) 119918
be found in the literature, it can be noted that the time resolution of The rest of the paper is described as follows: Section 2 defines
these optimization algorithms tend to be several minutes or longer, the properties of the employed RTS as well as the general equations
1 h being the most common assumption. While it is understandable and constraints applied in the system. Then, in Section 3, the pro-
that large time steps are desired to reduce the computational in- posed forecast-based optimization algorithm is described together
tensity when running these simulations, its direct drawback is that with two traditional benchmark algorithms. Section 4 describes the
the power values, such as solar power, are usually not constant case study, indicating the real and forecasted load, irradiance pro-
during this time range. Consequently, even when the optimization files, as well as the properties of the PV hybrid system. Subse-
results show an enhancement in the performance of their studied quently, Section 5 shows and discusses the results. Finally, the
systems, their potential for implementation in real systems remains overall conclusions of this work are provided in Section 6.
unknown. Hence, any proposed dispatch strategy algorithm needs
to be validated before it is implemented in a real hybrid system.
2. System model
Due to the large investment required to build a real system, another
option would be to employ real-time simulators (RTS), an alter-
The employed RTS is Typhoon Hardware in Loop 604 (HIL604).
native which can assess the system operation with a high level of
To carry out its modelings, the time is discretized by applying a
accuracy. Therefore, in this work we employ an RTS which is able to
time-slotted system. This is indexed by t, 1 t T, where T is the
perform simulations with a time step in the order of microseconds
total number of time slots. Dt, the slot size, which can be set in the
to test our optimization algorithms.
range of 0.5e20 ms, is set to 2 ms in this work. Fig. 1 a) presents the
Another limitation from typical scheduling optimization works
discussed hardware as well as the Typhoon SCADA interface for a
presented in the literature is the lack of analysis on how their al-
diesel generator. Fig. 1 b) shows the PV hybrid system in the
gorithms will affect the grid quality, namely, grid voltage, fre-
Typhoon Schematic Editor. The “PV system” block contains the
quency, and harmonics [7,8]. For example, when there is a high
solar array as well as the three-phase full-bridge PV inverter and
generation from renewable sources which goes beyond the load
filters. Similarly, the “Battery system” block contains the battery
demand, the grid voltage and frequency will tend to increase and
bank together with its three-phase full-bridge converter and
vice versa [9e11]. In addition, when high variations in the power
required filters. The “Diesel generator” block contains the diesel
demand (or generation) occur, transients will be produced [12].
generators (DGs) with their voltage/frequency controller and syn-
While is true that papers proposing scheduling optimization algo-
chronism blocks to assure that the diesel generators are connected
rithms have considered grid voltage variations, e.g. Refs. [13,14], the
to the grid only if they are in synchronism with it. The “load” block
frequency and harmonics estimation tends to be missing due to
contains the overall load demand of the system. Finally, the
their higher complexity. Transient analyses also tend to be avoided
“scheduling algorithm” block corresponds to the main controller
as transients tend to occur within a shorter time frame. As such, to
which generates the set points of the power generation and oper-
have a proper idea regarding the potential implementation of any
ation state of the energy sources of interest based on the applied
proposed scheduling algorithm in real life, its impact on the grid
algorithm (see Section 3).
should also be comprehensively analyzed. This has been taken into
The overall cost of the system can be split into the battery and
account in this work, where the grid voltage, frequency and har-
DG operational cost (CBAT and CDG , respectively):
monics have also been studied.
Furthermore, while most works only apply the present condi- X
T
tions, e.g. current irradiance and load demand, to run their pro- CT ¼ CBATðtÞ þ CDGðtÞ : (1)
posed scheduling algorithms, there is interest to perform t¼1
scheduling optimization while also considering the future weather
Because batteries degrade over time, they have a finite lifetime
and load conditions. Although there are works which aim to also
after which they need to be replaced. The charging/discharging
consider future weather/load conditions in their optimization code
lifetime, i.e. lifetime estimation based on the calculation of the total
[15e17], in general they do not apply a forecasting algorithm but
energy that a single battery will discharge during its lifetime,
tend to simply assume uncertainties to represent them. In our
E ½kWh [24], is considered in this work. The variable CBATðtÞ then
study, we propose to apply a real forecast approach on the load and
represents the cost of the battery to deliver energy, i.e. the more
weather-related profiles to properly analyze whether forecasting
energy it discharges, the sooner it needs to be replaced [25]:
can significantly enhance the scheduling optimization.
As PV hybrid systems have been installed in real life, real ap-
S,cBAT PdðtÞ ,Dt
proaches to control the dispatch strategy and ON/OFF state of DGs CBATðtÞ ¼ , ; (2)
have already been implemented. Two of them, referred in this work
E hBAT;inv
as the diesel generator backup (DGA) [18e21] and spinning reserve
where S ½kWh is the battery capacity and cBAT ½USD =kWh is the
(SRA) algorithms [22,23] are also implemented as benchmarks for
our proposed algorithm. In this study, we propose a forecast-based battery acquisition cost per unit energy, PdðtÞ ½kW corresponds to
algorithm (FBA) which aims to apply load and irradiance forecast to the discharging power from the battery inverters at time slot t and
better optimize the scheduling of the different energy sources in hBAT;inv ½% is the average weighted efficiency of these inverters.
real time. Hence, the main contributions of this work are: The cost associated to the diesel generators, CDGðtÞ [USD], is
calculated by estimating the amount of diesel used at time slot t,
1. A novel forecast-based approach has been proposed to optimize namely, kdieselðtÞ ½l. The diesel consumption is calculated as [26,27]:
the operation of a PV hybrid system in real time,
2. A real-time simulator has been employed to test the proposed nDG
X
and benchmark algorithms to accurately analyze their perfor- kdieselðtÞ ¼ aDG , PDGðj;tÞ þ bDG , PDG;nominalðjÞ , uðj;tÞ , Dt;
j¼1
mance and implementation potential in real life, and
3. The influence of the scheduling algorithms on the grid quality (3)
has been comprehensively analyzed, i.e. frequency, voltage and
harmonics. where PDGðj;tÞ ½W is the power produced by the jth DG at the time
slot t while uðj;tÞ indicates its status, i.e. uðj;tÞ ¼ 0 if DG is OFF, and
2
C.D. Rodríguez-Gallegos, L. Vinayagam, O. Gandhi et al. Energy 222 (2021) 119918
Fig. 1. a) Typhoon HIL604 hardware and SCADA interface; b) PV hybrid system in the Typhoon Schematic Editor.
uðj;tÞ ¼ 1 otherwise. The values of the coefficients aDG ½l =kWh and 2. Power produced by the battery bank: as the PV array operates at
bDG ½l =kWh depend on the quality of the DGs (the lower they are, its maximum power point with a power injected into the grid of
the less diesel is consumed to generate a specific power). The cost PPVðtÞ , and the “scheduling algorithm” block sets the desired
related to the operation of DGs can then be calculated: power output from the diesel generators, the power of the
battery system, in order to control the grid voltage and fre-
CDGðtÞ ¼ cdiesel ,kdieselðtÞ ; (4) quency, will be the one to assure that the power generated is
equal to the power demanded by the load PloadðtÞ . By neglecting
where cdiesel ½USD =l is the diesel price. the distribution losses, this relation can be presented as (see
In addition, there are also limitations on the power from bat- Fig. 1b)):
teries and diesel generators. The charging and discharging power of
the batteries at the AC side (PcðtÞ and PdðtÞ , respectively) need to be X
nDG
controlled to avoid the acceleration of degradation processes [28]: PPVðtÞ þ PdðtÞ PcðtÞ þ PDGðj;tÞ ,uðj;tÞ ¼ PloadðtÞ ;
j¼1
(8)
Pc;max ct2f1; /; Tg:
0 PcðtÞ ; ct2f1; /; Tg; (5)
hBAT;inv
Therefore, by knowing the power from the PV system, load
0 PdðtÞ Pd;max , hBAT;inv ; ct2f1; /; Tg; (6) demand, and by controlling the power from the diesel genera-
tors, the power from the battery system can be indirectly
where Pc;max ½kW and Pd;max ½kW are the maximum charging and controlled to a certain extent.
discharging power of the batteries, respectively. The three scheduling algorithms have three stages in their
To avoid issues related to DGs’ proper operation (such as oiling operation (shown in Fig. 2), and are ordered based on their
up of the silencer, bore glazing, and high temperature), their power priority:
is typically limited within their nominal ðPDG;nominalðjÞ Þ and a min-
imum value (PDG;minðjÞ ) [29]: 1. Battery safety,
2. DG ON/OFF state: define the ON/OFF state of diesel generators
PDG;minðjÞ PDGðj;tÞ PDG;nominalðjÞ ; under normal operation, i.e. when the battery safety stage has
(7)
ct2f1; /; Tg; j2f1; /; nDG g; uðj;tÞ 2f1g: not been activated, and
3. DG power: define the set points of the power of diesel genera-
tors under normal operation.
2. If the power produced from the PV system is lower than the load
demand, i.e. PPVðtÞ < PloadðtÞ , batteries and/or diesel generators
must also generate power based on the following criteria:
1. If the cost per kWh to discharge the batteries is lower than
the one from the diesel generators, i.e. S,cEBAT < aDG ,cdiesel , the
battery, together with the PV system, are used to satisfy the
load demand. If batteries cannot provide all the required
power, the DGs which are ON will then be employed as well.
2. If the cost per kWh to discharge the batteries is higher than
the one from the diesel generators, i.e. S,cEBAT > aDG ,cdiesel , the
DGs which are ON, together with the PV system, are used to
satisfy the load demand. If DGs cannot provide all the
required power, batteries will then be employed as well.
3. If the power from a DG is below PDG;min, it will then be set to
PDG;min to fulfill condition from Eq. (7).
Fig. 4. Power profiles, DG ON/OF state, battery SOC, and hourly cost results obtained from the DG-backup-based approach; spinning-reserve-based approach, and forecast-based
approach.
therefore, DG2 can be turned off as DG1 and the batteries can still
satisfy the load demand and fulfill the SR constraint. This operation
cycle continues to repeat until 8:21 AM. Notice that until this time,
in comparison to the DGA, the SRA is able to reduce the effective
number of operating hours of the DGs (which brings a cost
advantage). After 8:21 AM, it can be seen that now the two DGs will
be operating till the end of the day. This behavior is because by 8:21
AM, the load has increased to a value in which the SR condition
cannot be satisfied by a single diesel generator and therefore, the
two of them need to be operating. The reason why neither of the
diesel generators is turned off towards the end of the day (where
the load demand is decreasing), is due to the fact that the required
SR condition with the updated coefficients (εload;2 and εPV;2 ) is not
able to be fulfilled if one of these generators is turned off.
Fig. 5. Cumulative cost for the analyzed case study obtained from the benchmark and
Finally, when analyzing the results from the proposed FBA al-
proposed algorithms.
gorithm (see Fig. 4g) and h) and i)), it can be noted that their
behavior is quite similar to the ones from the SRA. Nevertheless,
after 8:21 AM, it is not necessary for the two DGs to be continuously from this algorithm becomes the highest after a short time. It can
operating. The reason is that, based on the forecasting results, it is also be noted that the total costs incurred by SRA and FBA are
expected that a single diesel generator together with the batteries similar until 8:21 AM as the power profiles and the operation states
can satisfy the expected load demand and as such, DG2 will only be of their DGs are the same, as previously explained. However, after
activated when the batteries’ SOC reach their minimum and will be this point, the SRA approach turns ON the second DG and, although
disconnected when the SOCbsmin value is achieved. As a result, by the power profiles are still similar, the SRA approach starts to incur
taking into account the forecasted irradiance and load profiles, a higher cost than FBA. This is a consequence of the diesel con-
better decisions were taken which generated lower costs during sumption required to keep a DG ON (see Eq. (3)) even while
multiple operating hours, i.e. when comparing Fig. 4 f) and i), generating the same total power. For example, although the power
although both incur similar costs in the morning, the latter ach- required to be produced by the DGs is similar at 11:00 AM, there is
ieves, in general, lower hourly costs after 8 AM. only one operating DG when using the FBA approach (see Fig. 4 h))
Fig. 5 presents a closer view to the overall costs. It can be noticed while the two DGs are operating in the SRA approach (see Fig. 4 e)).
that, as the DGA approach employed only the batteries until 0:56 Therefore, the latter will consume extra diesel to keep the second
AM, while the other algorithms also operated their DG (see Fig. 4 b), DG in ON state. This shows the advantage of FBA in determining the
e), h)), the former produces the lowest cost until that point in time. ON/OFF state of DGs to reduce the diesel consumption and conse-
However, after 0:56 AM, the battery SOC for the DGA approach has quently the overall cost. For all the algorithms, batteries generated
reached its minimum, therefore, the two DGs start to operate. the lowest cost. In addition, the total cost at the end of the day was
Consequently, due to the high diesel consumption, the overall cost similar when using the benchmark algorithms (5967 USD for DGA
6
C.D. Rodríguez-Gallegos, L. Vinayagam, O. Gandhi et al. Energy 222 (2021) 119918
Fig. 6. Simulated grid parameters, i.e. frequency, line-line voltage and their total harmonic distortion, from applying the benchmark and proposed algorithms in RTS.
and 5989 USD for SRA) while the minimum cost was achieved with analyzed optimization algorithms and the results showed that the
the proposed FBA (5678 USD), generating 5% savings. frequency, voltage, and total harmonic distortion were kept within
With respect to the grid quality, Fig. 6 presents the grid fre- the allowed limits. This outcome suggests that the proposed algo-
quency, voltage, and the total voltage harmonic distortion, rithm could be applied in a real system without seriously affecting
observed for all the approaches. The safety ranges defined in Sec- the grid quality.
tion 4 have not been violated, giving more confidence to apply the Based on this work it can be concluded that the application of
proposed algorithm in a real application. forecasting approaches can be beneficial to optimize the perfor-
While this work successfully showed that introducing fore- mance of PV hybrid systems. Nevertheless, special attention should
casting approaches has the potential to enhance the scheduling be paid to the forecasting accuracy as large forecasting errors could
optimization of PV hybrid systems, it can be understood that its reduce the system performance. Hence, more work is necessary to
applicability can be extended to other system configurations as further enhance the accuracy of the forecasting methods as well as
long as there are parameters that can be forecasted. to perform this scheduling optimization for bigger systems
considering a larger time range to fully validate the effectiveness of
6. Conclusion the proposed approach.
In this work, a scheduling optimization algorithm was proposed Declaration of competing interest
to optimize the operation of off-grid systems composed of solar
panels, batteries and diesel generators. This algorithm incorporates The authors declare that they have no known competing
load and irradiance forecast to set the optimal power outputs of the financial interests or personal relationships that could have
energy sources and to control their operation state, e.g. when to appeared to influence the work reported in this paper.
turn ON/OFF the diesel generators. As benchmarks, two algorithms
which are typically used in real PV hybrid systems were consid- Acknowledgment
ered: 1) the DG-backup-based algorithm which only employs the
DGs when the battery has achieved a minimum state of charge; and This work was supported by the Singapore’s National Research
2) the spinning-reserve-based algorithm which evaluates whether Foundation Singapore within the research project “NRF2013EWT-
there is enough spinning reserve to decide on the ON/OFF state of EIRP003-043”. SERIS is supported by the National University of
the DGs. In an effort to validate the results generated, a real-time Singapore and Singapore’s National Research Foundation (NRF)
simulator, i.e. Typhoon Hardware in Loop 604 - HIL604, was used through the Singapore Economic Development Board (EDB).
to run the applied optimization algorithms on a resort case study.
For the analyzed case study, the results revealed that the pro- References
posed forecast-based algorithm is able to achieve the lowest cost,
yielding up to 5% cost savings. This outcome was produced because [1] Tang R, Wu Z, Li X. Optimal operation of photovoltaic/battery/diesel/cold-
the forecasting estimations were accurate enough to make good ironing hybrid energy system for maritime application. Energy 2018;162:
697e714.
decisions on the dispatch strategy and operation state of the energy [2] Tu T, Rajarathnam GP, Vassallo AM. Optimization of a stand-alone photo-
sources. In addition, the grid quality was evaluated for all the voltaicewindedieselebattery system with multi-layered demand scheduling.
7
C.D. Rodríguez-Gallegos, L. Vinayagam, O. Gandhi et al. Energy 222 (2021) 119918