Professional Documents
Culture Documents
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
H I G H L I G H T S
A R T I C L E I N F O A B S T R A C T
Keywords: The effective design of Distributed Energy Systems (DES) is subject to multiple uncertainties related to aspects
Distributed energy system like the availability of renewable energy, the building energy demands, and the energy carrier prices.
Energy hub optimization Nevertheless, current practices involve the use of deterministic design models, which overlook uncertainty and
Monte Carlo uncertainty analysis can lead to suboptimal DES configurations that fail to deliver the desired performance.
Global sensitivity analysis
A necessary condition in order to obtain robust DES designs against uncertainty is the understanding of
Morris screening
uncertainty’s impacts and main drivers. Therefore, this paper presents a novel methodological framework for the
Sobol indices
investigation of uncertainty in the context of DES design, which combines optimisation-based DES models and
techniques from Uncertainty Analysis (UA) and Global Sensitivity Analysis (GSA). Moreover, the application of
the framework is illustrated with a case study for the optimal DES design of a Swiss urban neighbourhood.
Embarking from a deterministic DES design model, first, all sources of uncertainty are identified and ap-
propriate probabilistic descriptions are assigned to all uncertain model parameters. UA is then performed using
Monte Carlo (MC) simulations to quantify the impacts of uncertainty. Results reveal substantial variations in
terms of economic and carbon performance of the optimal DES, but also in terms of optimal DES configurations.
Moreover, the UA results indicate that the optimal system costs are mostly higher than the deterministic cost
estimates, while the inverse is observed for the case of carbon emissions. Additionally, in many of the MC
simulations, the resulting optimal DES configurations deviate significantly from the deterministically obtained
designs, which confirms the risk of suboptimal decisions in deterministic design processes. Moreover, the results
of UA can function as decision support by identifying the DES configurations that are optimal for most rea-
lisations of uncertainty.
Finally, to identify the uncertain parameters that are mostly responsible for the variation of the economic
performance of the DES, a two-step GSA is launched, combining the Morris method and the variance-based Sobol
method. Results of the GSA indicate the energy demand patterns and the energy carrier prices as primarily
responsible for the variability of the optimal system cost, while parameters like the investment costs and the
technical characteristics of the technologies exert only minimal influence. The results of GSA, besides offering a
better understanding of uncertainty to DES designers, also identify the parameters for which additional effort
needs to be invested to reduce their uncertainty and, as a result, the uncertainty associated with DES design.
⁎
Corresponding author at: Chair of Building Physics, ETH Zurich, Stefano-Franscini-Platz 1, 8093 Zurich, Switzerland.
E-mail address: gmavroma@ethz.ch (G. Mavromatidis).
https://doi.org/10.1016/j.apenergy.2018.01.062
Received 7 October 2017; Received in revised form 18 December 2017; Accepted 22 January 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
220
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Fig. 1. General workflow for uncertainty investigation in DES (illustration inspired by Sudret et al. [52]).
the DES literature to integrate DES design models with methods of UA & network is used.
GSA and to showcase their application and full potential. Moreover, This section presents the main model concepts; the complete
another novelty of this paper is that by employing BPS tools to char- mathematical formulation is given in Appendix A. The model identifies
acterise the uncertainty of building energy demands, it is the first effort simultaneously the optimal DES design (technology selection and
to bridge the disciplines of uncertainty in BPS and DES design. sizing) and its optimal operating strategies. It is based on the energy
The paper is structured as follows: Section 2 details the metho- hub modelling framework [53] and it is a mixed-integer linear program
dology for uncertainty investigation including a deterministic DES de- (MILP) following the work done in previous papers [54–57].
sign model, Uncertainty Characterisation (UC), Uncertainty Analysis The minimisation of the total system cost and of the total CO2
(UA) and Global Sensitivity Analysis (GSA). Section 3 introduces the emissions are the objectives considered for the DES design. Following
paper’s case study for DES design. Section 4 presents and discusses the the ε-constraint method [58], the former objective forms the model’s
paper’s results regarding both the deterministic model and the UA and objective function, while the latter is expressed in the form of a con-
SA. Finally, Section 5 provides concluding remarks. straint setting an upper limit to the CO2 emissions, as shown in Eqs. (1)
and (2).
2. Methodology for uncertainty investigation min Total system cost: EAC = Invest + Op (1)
This paper’s methodology for uncertainty investigation in DES de- s. t . CO2 emissions ⩽ CO2 limit (2)
sign is illustrated in Fig. 1. Step A includes the development of a de-
The total system cost (Eq. (1)) is expressed as the Equivalent Annual
terministic optimisation model for optimal DES design. Step B includes
Cost (EAC) and is calculated for a typical representative year. The EAC
Uncertainty Characterisation (UC) in which the uncertain parameters of
consists of an annualised investment cost, Invest , which includes the
the deterministic model are identified and attributed probabilistic de-
investment for the generation, storage and distribution technologies of
scriptions. In Step C, Uncertainty Analysis (UA) is performed using
the DES. The term Op represents the total annual operating expenditure
Monte Carlo (MC) simulations. Finally, in Step D, a two-step Global
and includes the cost for the consumption of energy carriers and the
Sensitivity Analysis (GSA) identifies the most important uncertain
income stream associated with electricity exports. The CO2 emissions
parameters by calculating Sobol indices as importance measures. Steps
(Eq. (2)) are also calculated over the period of a year and account for
A to D are described in Sections 2.1–2.4 and a discussion of the
the system’s operating emissions and not for its life-cycle emissions. The
methodology’s strengths and criticisms is given in Section 2.5.
CO2 limit can be selected by the DES designer and varied to obtain
multiple designs that achieve different carbon performance. The defi-
2.1. Deterministic DES design model nitions of the terms in Eqs. (1) and (2) are given in Appendix A.
Besides Eq. (2), additional model constraints are required to describe
In the methodology’s first step, a deterministic optimisation model the system’s energy balances and other operational and technical con-
is developed for the task of optimal DES design. The candidate tech- straints. The formulation of all these constraints is found in Appendix A.
nologies for the DES are presented in Fig. 2 and include: a biomass As inputs to the model, parameters pertaining to the technical (e.g.
boiler (BIO), a gas boiler (GAS), a gas-fired CHP engine (CHP), a conversion efficiencies etc.), economic (e.g. investment costs, energy
ground-source heat pump (GSHP), a thermal storage tank (TS), a bat- carrier prices etc.) and environmental dimensions of the energy tech-
tery (BAT) and photovoltaic panels (PV). All technologies are con- nologies (e.g. emission factors) are required, in addition to time series
sidered for installation in the DES except for PV panels, which are to be profiles of building energy demands and solar radiation availability.
installed in a spatially distributed manner on the individual building The deterministic parameter values are given in Section A.2 of
roofs. The generated electricity is delivered to the buildings via the Appendix A.
local distribution grid, while for thermal energy a district heating Finally, model decision variables can be split into design variables that
221
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Fig. 2. Superstructure representation of candidate set of energy technologies for the DES.
represent the selection and the capacity of a generation/storage tech- uncertainty range for 2021 is set to ±2%, increasing at a 2% rate per
nology, and operating variables that represent the utilisation of those year.2 Overall, the cone reflects how uncertainty increases as the pro-
technologies and the energy imports/exports for each model time step. jection horizon moves further into the future. Moreover, as can be
observed in Fig. 3, the prices are mostly projected to increase compared
2.2. Uncertainty characterisation (UC) to current prices (see Appendix A).
As a final step, a uniform distribution is constructed per energy
Uncertainty Characterisation (UC) entails the identification of a carrier price by taking the 20-year means of the most and the least
model’s uncertain parameters and the assignment of a mathematical expensive price pathways (see Fig. 3b), similarly to Carpentiero et al.
description to their uncertainty. Probabilistic methods, which treat [66]. The distributions for all energy carriers are given in Table 1.
parameters as random variables following Probability Density Regarding the feed-in tariff for exported electricity, FiT, due to the lack
Functions (PDF), are the most common to describe parameter un- of any projections, a uniform distribution shown in Table 1 is assigned
certainty [59,60]. A review of UC approaches for uncertain parameters to it allowing a ± 30% variation around the nominal value.
in DES design models can be found in [61]. In the following sections,
the UC approaches for all the parameter categories of the DES design 2.2.2. Emission factors
model are discussed in detail. In total, 27 individual uncertain para- Additional uncertain parameters pertain to the energy carrier
meters are identified for the model of Section 2.1 and a summary table emission factors. The natural gas emission factor, however, is con-
is given in Appendix B. sidered constant due to negligible variations in the fuel’s carbon con-
tent. Additionally, the factors of renewable energy carriers are taken as
2.2.1. Energy carrier prices zero. Hence, only the emission factor associated with grid electricity is
For the energy carrier prices, deterministic models commonly use considered uncertain due to the dynamic character of a country’s
values that correspond to current prices at the time of design (see e.g. electricity mix that evolves driven by environmental regulations, en-
[10,13,48]). Nevertheless, during the long lifetime of a DES, prices are ergy targets etc.
expected to vary driven by a multitude of factors like energy markets, Currently, the Swiss grid emission factor for is very low (9.5 gCO2/
energy policy etc. [62]. Their exact evolution is difficult to predict, kWh, see Appendix A) due to the high shares of nuclear and hydro
hence, they should be treated as uncertain. In this work, to characterise energy in its generation mix [67]. However, Switzerland’s decision to
the uncertainty of energy carrier prices information is sourced from the phase-out nuclear energy by 2033 [67] is expected to alter the country’s
Swiss Energy Strategy 2050 (SES2050) [63]. Projections of future nat- electricity generation map. The degree and the direction of change
ural gas, biomass, and electricity prices are taken for the period cannot be known in advance and, thus, the SES2050 offers multiple
2021–2040, which is selected as the system’s operational horizon. projections for the country’s electricity demands and electricity supply
Three scenario projections are available (Weiter Wie Bisher – WWB, [63].
Politische Massnahmen – POM, Neue Energiepolitik – NEP),1 which Thus, to characterise the uncertainty of the Swiss grid emission
differ in terms of the Swiss energy landscape they describe [63,64]. The factor, a similar approach as with the energy carrier prices is adopted.
electricity price projections are illustrated in Fig. 3a. For all electricity demand and supply scenarios of SES2050, the evo-
However, these scenarios represent only two possible futures and lution of the grid emission factor is calculated and used to construct a
not the full array of possibilities (see e.g. analogous projections in [64]). uniform distribution by taking the minimum and the maximum values.
To capture a wider range of possible prices, a “cone of uncertainty” is The resulting range of the distribution is [10.1–85.7] gCO2/kWh.
assigned to each SES2050 price projection. The cone is constructed as
follows: Taking e.g. the electricity prices, each scenario’s price for 2021
2.2.3. Investment costs
is assigned a ±1% uncertainty range. Similarly, for 2022, a ±2% range
The investment costs for the generation, storage and distribution
is assigned and so on until a ±20% range is assigned to the prices of
2040 as can be seen in Fig. 3b. The same process is applied for biomass
prices too, while due to the higher price volatility of gas prices, the 2
The choice to attribute the 2% increase rate value for the uncertainty ranges of the
fossil fuel carrier prices was also made by examining the evolution of the consumer prices
over the period between 2000 and 2015, as reported by the Swiss Federal Office of Energy
1
Note that the price projections for the WWB and the POM scenarios in the SES2050 (SFOE) [65], which reveal much higher fluctuations for the natural gas prices compared
are the same. to electricity.
222
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
40 40
a b
25 25 Least
expensive
pathway
2021 2025 2030 2035 2040 2021 2025 2030 2035 2040
Years Years
Fig. 3. Electricity price evolution according to SES2050 and associated uncertainty cones.
223
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
2.3. Uncertainty Analysis (UA) quantify how much more important one parameter is over another.
Nevertheless, this does not set a limitation, as another GSA technique is
To perform UA, a deterministic model y = g (x ) of the examined applied in the second step on the set of parameters that are not elimi-
system (Section 2.1) and a probabilistic framework for its uncertain nated by the Morris method to provide more quantitative information
parameters (Section 2.2) are required. The first step then involves the about parameter importance.
generation of N samples for each of the k uncertain input parameters This GSA method is the variance-based Sobol method [85]. Ad-
using their probability distributions. These samples are stored in a vantages of this method include that no assumptions on the model form
matrix X (Eq. (3)), in which x i,j represents the value of the jth sample of are required (e.g. linearity or monotonicity), it covers the full range of
a parameter Xi . The sampling from parameter distributions is per- input parameters considering their distributions, and it provides quan-
formed using the Sobol sequence for its better space filling properties titative information for parameter importance. In the Sobol method, the
compared to random sampling [84]. model output variance, V (Y ) , is used as a measure of its uncertainty.
x x ⋯ xk,1 The method’s goal is to quantify the contribution of each uncertain input
⎡ 1,1 2,1 ⎤ parameter to the output’s variance by calculating two metrics per para-
x1,2 x2,2 ⋯ xk,2 ⎥
X=⎢ meter named first-order Sobol index, Si , and total-order Sobol index, STi
⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ x1,N x2,N ⋯ xk,N ⎥ (3) (see Eqs. (C.3) and (C.4)). These indices quantify the portion of the output
⎣ ⎦
variance V (Y ) that is explained by a parameter alone (first-order index)
An important aspect during sample generation is the treatment of and by a parameter and its interaction effects5 with other parameters
possible parameter correlations. In this paper, no correlations are taken (total-order index). Hence, high Sobol index values indicate that a
between investment costs, since the uncertainty in estimating one parameter is highly significant regarding its influence on the output
technology’s price (e.g. a heat pump) does not influence the estimate for variations.
another technology’s price (e.g. a boiler). The same argument applies to Methods for calculating these indices are based on MC simulations,
the devices’ technical characteristics; hence, no correlations are taken. which is the main weakness of this technique [34] as state-of-the-art
Regarding the energy demand and solar radiation patterns, their cor- methods require n(k + 2) model evaluations, with k being the number
relations are treated inherently by the approach in Section 2.2.5, which of uncertain parameters and n a parameter with a suggested value of
calculates them with a BPS tool and compounds them in a discrete 500–1000 [35]. Thus, while the Sobol method can be applied directly to
random variable. Finally, regarding energy carrier prices, the prices of the full parameter set, if the number of uncertain parameters is large,
biomass, natural gas and electricity over the last two decades were the computational cost can be prohibitive. Hence, the use of the Morris
analysed and no significant correlations were found. The diagrams method as a first screening step is recommended [51].
showing the price variations and their weak correlations are given in
the paper’s Supplementary Material.
2.5. Method discussion and critical reflections
Once the sample matrix is generated, the last step involves N Monte
Carlo evaluations of the deterministic model - one per row of the matrix
UA and GSA offer quantitative insights about uncertainty’s effects
- to obtain the variable model response, y , (see Eq. (4)). By evaluating
on a model’s output and they indicate the parameters from which un-
the variability and the statistical characteristics of the model’s variable
certainty mainly stems. Additionally, since they rely on Monte Carlo
output the issue of quantifying uncertainty’s impacts is addressed.
simulations of a deterministic model, they do not require any mod-
y g (x1,1,x2,1,…,xk,1) ⎤ ification to the models that are already used in deterministic DES design
⎡ 1⎤ ⎡ making their adoption easier.
⎢ y2 ⎥ ⎢ g (x1,2,x2,2,…,xk,2) ⎥
y= =
⎢⋮⎥ ⎢ ⋮
⎥ However, UA & GSA are decision-support techniques, rather than
⎢ yN ⎥ ⎢ ⎥
decision-making techniques. While UA results can assist in deciding on
⎣ ⎦ ⎢ ⎣ g (x1,N ,x2,N ,…,xk,N ) ⎥
⎦ (4)
an optimal DES configuration (e.g. the one that emerges most often as
optimal), they cannot point towards a single decision that is optimal for a
2.4. Sensitivity Analysis (SA) specific criterion against uncertainty. This is instead the task of tech-
niques for Optimisation under Uncertainty (OU2) [84] (e.g. Stochastic
The final step of the methodology is to perform SA to identify the Programming (SP) [85]), which incorporate uncertainty directly in the
most influential uncertain parameters regarding the observed output model and allow decisions to be made before uncertainty is resolved.
variations. Given the drawbacks of LSA, a two-step GSA approach is Nevertheless, UA, GSA and OU2 are not competing but rather sy-
used, as shown in Fig. 4. The technical details of the methods are given nergetic techniques. It is useful to confirm whether uncertainty has a
in Appendix C, but their main concepts are briefly described here as significant impact on DES design via performing UA, before a more
well. More extensive information on the methods used can be found in complex model based on SP is developed. Moreover, GSA can benefit
[35]. model development by identifying the most important uncertain para-
In the first step, the Morris method [51] is applied on the full set of k meters and the ones whose uncertainty can be safely ignored.
uncertain parameters to screen out unimportant ones at a low compu-
tational cost as it requires r(k + 1) MC model evaluations (re- 3. Case study
commended values for r are between 4 and 10) [35]. The method of
Morris ranks parameter importance by calculating a metric μ∗ per un- 3.1. Description
certain parameter (Eq. (C.2)), which quantifies its overall influence on
the output. However, the ranking is only qualitative i.e. it cannot As a case study for this paper, the design of a DES for a fictitious
urban neighbourhood in Zurich, Switzerland, is investigated. The
neighbourhood, presented in Fig. 5a and b in 2D and 3D representa-
tions, is composed of ten buildings of different types and ages. Multi-
family houses (MFH) and offices are included, which are also combined
with other commercial premises to form the MFH-mixed and Office-
Fig. 4. Two-step GSA procedure consisting of a screening first step using the Morris
method and a second step using the variance-based Sobol method on a smaller subset of 5
Two parameters are interacting when their combined effect is different than the sum
the most influential uncertain parameters. of their single effects on the output.
224
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Fig. 5. a. Building age and type of the Swiss urban neighbourhood case study, and neighbourhood DES location, b. 3D illustration of the urban neighbourhood case study.
mixed categories. A centralised neighbourhood DES is envisioned that range. Hence, when random values are sampled from the parameters’
generates, stores and distributes energy to the individual buildings and distributions in the stochastic case, it is highly likely these will be
PV panels are to be installed on the buildings’ roofs. The location of the smaller than the deterministic ones, which in turn lead to smaller
neighbourhood DES and the district heating network’s path are also electricity demands. Overall, though, significant energy demand var-
illustrated in Fig. 5a. iations are observed, which the deterministic profiles cannot reflect. For
a better understanding of the seasonal distribution of heat demands,
Fig. 6b presents the monthly energy demands for the deterministic and
3.2. Deterministic and variable energy demands
all the variable profiles.
Finally, as was mentioned in Section 2.2.5, in order to be able to
In this section, the energy demands of the case study buildings are
sample from the generated set of profiles, a proxy variable Ltotal is de-
presented – both deterministic and stochastic. The former are generated
fined that follows a discrete uniform distribution, defined as U {1,1440} .
by running an EnergyPlus simulation per building using nominal values
for all input parameters and a weather file reflecting today’s climate
4. Results
(more information is given in Section A.2). For the variable profiles, the
process from Section 2.2.5 is followed and 1440 energy demand and
4.1. Deterministic results
solar radiation profiles are generated per building. The demands in both
cases are aggregated to form the neighbourhood energy demands.
Initially, the results of the deterministic analysis are presented. In
The results are presented in terms of annual heat and electricity
total, five scenarios are considered in which cost-optimal DES are
demands with the violin plots in Fig. 6a. Since the future, warmer cli-
sought with different emission requirements set according to Eq. (2).
mate is considered for the stochastic case, the neighbourhood’s heat
The scenarios are:
demands are projected to mostly decrease compared to the determi-
• Scenario 1: No CO limit.
nistic value. Additionally, the electricity demands are also mostly
2
120 500
a b
Deterministic profile
Variable profile
100 400
demand (kWh/m2 )
Annual energy
demand (MWh)
Monthly heat
80 300
200
60 Determ. point
100
40
Heat Electricity 0
Jan Mar May Jul Sep Nov
demand demand
Fig. 6. a. Violin plots for the annual neighbourhood heat and electricity demands, b. Variable monthly neighbourhood heat demands.
225
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
30
a Scenarios c
0
100
No carbon limit
10
Carbon emissions (kgCO 2 /m2 ) 90
25 20 kgCO2 /m 2
20
15 kgCO2 /m 2 80
30
10 kgCO2 /m 2 70
%)
40
20
Gr
2 60
5 kgCO 2 /m
il (
id
50
ss
(%
50
Fo
)
60
15 40
70
30
80
20
10
90
10
10
0
0
0
10
20
30
40
50
60
70
80
90
0
5
10
25.5 26 26.5 27 27.5 Renewable (%)
2
Total cost (CHF/m )
1200
b
Technologies
1000 Oil boiler
Bio boiler
Capacity (kW | kWh | m 2 )
CHP
800 GSHP
Th. Storage
PV
600
400
200
0
No carbon limit 20 kgCO2 /m2 15 kgCO2 /m2 10 kgCO2 /m2 5 kgCO2 /m2
Scenarios
Fig. 7. a. Total system cost and carbon emissions of the deterministic optimal configurations at the different scenarios, b. Technology selection and capacities for the deterministic optimal
system configurations for the different scenarios, c. Energy consumption shares for the deterministic optimal configurations of the different scenarios.
The optimal systems’ cost and carbon performance are given in is eliminated, while in Scenario 5, a biomass boiler is added to achieve
Fig. 7a normalised per m2 of total building floor area. In the first sce- emissions below 5 kgCO2/m2. Finally, the thermal storage capacity is
nario, the system’s total cost is approx. 25.5 CHF/m2, while its emis- the same across all scenarios indicating its high value.
sions are equal to 25 kgCO2/m2. Using the latter value, the CO2 limit Besides design aspects, the model of Section 2 also calculates the
values of Eq. (2) for Scenarios 2 to 5 become 20, 15, 10, and 5 kgCO2/ optimal operating patterns of each technology. Thus, the percentage
m2, respectively. In these scenarios, as more stringent emission con- shares of grid electricity, fossil fuels (natural gas) and renewable energy
straints are imposed, the total cost is increased. However, the rate of (biomass and PV electricity) in the system’s total energy consumption
increase remains small, with the total cost in Scenario 5 (27.1 CHF/m2) (TEC) can be calculated. The mathematical definitions of those terms
being only 6.3% higher compared to the cost of Scenario 1. are given in Appendix A. The energy shares for the different scenarios
The corresponding optimal DES configurations are given in Fig. 7b. are plotted in Fig. 7c using ternary plots.
For Scenario 1, a three-generation-system configuration emerges com- In Scenario 1, the system relies heavily on natural gas, which cor-
posed of a GSHP, a gas boiler and a CHP engine, while the PV area is responds to almost 90% of the TEC, with the grid and renewable shares
equal to 816 m2 (65% of the total roof area). Due to emission con- being below 10% each. This is attributed to the high CHP engine uti-
straints in Scenarios 2 to 5, the CHP capacity is gradually reduced and lisation that covers large parts of the neighbourhood’s heat and elec-
the gas boiler and GSHP capacities are increased. In Scenario 4, the CHP tricity demands. From Scenario 2 to Scenario 4 the fossil share is
226
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
No carbon limit
20 kgCO2 /m 2
Scenarios
15 kgCO2 /m 2
10 kgCO2 /m 2
5 kgCO 2 /m 2
Determ. value
25 30 35 0 10 20 30
2
Total cost (CHF/m ) Carbon emissions (kgCO 2 /m 2 )
Fig. 8. Violin plots for the variation of the total system cost and the carbon emissions for the different scenarios.
progressively reduced, while the grid share is increased indicating emerge. Hence, gas boilers, GSHPs, thermal storage and photovoltaics
higher GSHP utilisation and, hence, lower emissions. This also explains are popular technologies for the configurations of all scenarios. On the
why the total emissions are reduced despite the increasing gas boiler other hand, CHP engines are included in many configurations of
capacities in Fig. 7b. Finally, while the renewable share remains rela- Scenarios 1 and 2, but from Scenario 3 onwards their popularity is
tively constant in Scenarios 2 to 4, the addition of the biomass boiler in decreasing until Scenarios 4 and 5 where they are hardly ever installed.
Scenario 5 pushes it to almost 30%, while the fossil share is limited to The same visualisation also illustrates the gradual increase of biomass
approx. 20%. boiler selection for the optimal configurations. Unlike the deterministic
case when a biomass boiler was only included in Scenario 5, when
uncertainty is considered, these devices become “popular” already from
4.2. Uncertainty Analysis results
Scenario 4. Such results can offer decision support to decision-makers
regarding the selection of an optimal DES configuration by selecting,
For the UA step of the methodology, 5000 MC evaluations of the
for instance, the one that emerges most often as the optimal one.
deterministic model of Section 2.1 are performed. The five scenarios of
Another important dimension pertains to the fact that many optimal
the deterministic case and their CO2 limit values are maintained to
configurations from the MC simulations deviate from the deterministic
contrast the UA results against the deterministic results.
solutions. To get a quantitative perspective on this matter, Fig. 10
Fig. 8 illustrates the variation of the systems’ cost and carbon per-
presents the percentage of MC runs per scenario for which such de-
formance using violin plots. Starting from the cost distributions, sig-
viations are observed in terms of technology selection (technology ca-
nificant variations are observed across all scenarios along with a gen-
pacity differences are not considered). This percentage essentially then
eral shift to higher values for lower emission requirements. Moreover,
represents the probability that the deterministic configurations are ac-
in most cases the costs are considerably higher than the deterministic
tually suboptimal decisions.
values. Therefore, deterministic cost estimates can be considered un-
For Scenarios 1 to 3, significant portions of the MC runs, ranging
reliable, which is an important issue for stakeholders in DES decision-
from 47% to 75%, result in different solutions. These deviations are
making.
primarily due to many MC runs not including a CHP engine in their
Regarding CO2 emissions, in Scenario 1, the distribution’s median is
configuration unlike their deterministic solutions. Additionally, the
lower than the deterministic value, mainly due to lower natural gas
tighter the emission constraint, the more likely it is for a CHP engine
consumption, which in turn is due to its projected increasing prices
not to be included in the configurations; hence, the MC run percentage
(Section 2.2.1). Additionally, the violin’s form reveals a bimodal
that deviate from the deterministic configuration increases. For
emission distribution, which is due to the large disparities between the
Scenario 4, the percentage remains high and the discrepancy between
carriers’ emission factors. Systems that rely more on fossil fuels cluster
the MC and the deterministic solutions is mainly driven by biomass
at the higher emissions side and systems that rely more on grid elec-
boilers. In the deterministic case, such a boiler type was not included,
tricity and renewables cluster at the lower part. Additionally, despite
while for many of the MC runs, as Fig. 9 shows, a biomass boiler is
the lack of an emission constraint in Scenario 1, there are cases that
necessary to achieve the required carbon performance. Finally, for
result in lower emissions than the limits in Scenarios 2 to 5. In Scenarios
scenario 5, the deterministic system selection seems rather robust, since
2 to 5, in which, emission limits are set, the emission distributions show
as it is also evidenced by Fig. 9, most configurations relying on a GSHP
most points being concentrated at each scenario’s emission limit, while
and a biomass boiler are aligned with the deterministic design.
there also instances where the actual emissions are below the limit.
Finally, Fig. 11 presents the variation of the systems’ energy shares
Additional insight is gained by investigating the variations of the
and contrasts them against their corresponding deterministic points. A
optimal system configurations with the “parallel coordinate” plots of
first observation pertains to the considerable variation that the shares
Fig. 9. In this plot, the optimal design from each MC run is visualised
exhibit across all scenarios. For Scenario 1, for instance, there are cases
with a polyline intersecting vertical axes corresponding to energy
where the grid share is almost 0% and others where it reaches 70%, the
conversion and storage devices. The intersection points between the
renewable share can be above 40% and almost 70%, and there are other
axes and a polyline indicate the device’s capacity (if a device is not
cases where the fossil shares exceed 90%. In Scenarios 2 and 3, similar
selected in any MC run, it is completely omitted from the plot).
patterns emerge along with a general shift of the points towards
This visualization allows the technology selection patterns to
227
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
No carbon limit
1257 kW 250 kW 370 kWth 1024 kW 600 kWh 1246 m2
Capacity range
10 kgCO2 /m 2
Capacity range
Gas B. GSHP CHP Bio B. Th. Stor. PV
System capacities
20 kgCO2 /m 2
System capacities
5 kgCO2 /m 2
Capacity range
Gas B. GSHP CHP Bio B. Th. Stor. PV
System capacities
15 kgCO2 /m 2
System capacities
to the fossil fuel usage. However, in the uncertainty domain and given
the unknown future Swiss grid carbon factor, a higher reliance on grid
electricity and renewable energy might be required to meet the re-
quired emission levels.
228
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
0
100 100
10
10
90 90
20
20
80 80
30
30
Gr
70 70
)
l (%
l (%
40
40
id
Gr
60 60
(%
id
ssi
ssi
50
50
(%
)
Fo
Fo
50 50
)
60
60
40 40
70
70
30 30
80
80
20 20
90
90
MC point MC point
10 10
10
10
Determ. point Determ. point
0
0
0 0
0
10
20
30
40
50
60
70
80
90
0
0
10
20
30
40
50
60
70
80
90
0
10
10
Renewable (%) Renewable (%)
0
100 100
10
10
90 90
20
20
80 80
30
30
70 70
)
)
Gr
Gr
l (%
l (%
40
40
id
id
60 60
ssi
ssi
50
50
(%
(%
50 50
Fo
Fo
)
)
60
60
40 40
70
70
30 30
80
80
20 20
90
90
MC point MC point
10 10
10
10
0 0
0
10
20
30
40
50
60
70
80
90
0
0
10
20
30
40
50
60
70
80
90
0
10
10
Renewable (%) Renewable (%)
5 kgCO2 /m2
0
100
10
90
20
80
30
70
Gr
)
l (%
40
id
60
(%
ssi
50
50
)
Fo
60
40
70
30
80
20
90
MC point
10
10
Determ. point
0
0
0
10
20
30
40
50
60
70
80
90
0
10
Renewable (%)
Fig. 11. Grid electricity, fossil fuel and renewable energy share variations for the different scenarios.
calculated by applying the Sobol method on this screened parameter set parameters, the Sobol indices reveal that rather a small set of these
are also reported in the figure. A first observation pertains to the fact parameters is responsible for the greatest portion of the output’s var-
that the Morris results are aligned with the Sobol results in terms of iance, indicated by the sum of the first- and total-order indices. In all
parameter ranking as the Sobol indices decrease in the same order as cases, this set is composed by the energy demands/solar radiation
the value for μ∗. Therefore, if only a quantitative ranking was required, (Ltotal) and the energy carrier prices (OCm), but at different relative
GSA could be performed using only the Morris method. This could also importance. In Scenario 1, due to the absence of any emission con-
limit the number of MC model runs required. straints, systems that rely heavily on natural gas emerge; hence, the gas
Even though the Morris method has identified the most important price, OCgas, is the most important parameter. Second most important
229
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Sobol indices
Sobol indices
Sobol indices
0.4
6 0.3 0.3
*
*
1 0.3 1
4 0.2 0.2
0.2
0 0 0 0 0 0
nnet
nnet
nnet
nCHP
Inv CHP
nGAS
nCHP
nGAS
Inv CHP
nGAS
Inv GAS
Ltotal
Inv GSHP
Inv PV
COP GSHP
Ltotal
Inv GSHP
Inv PV
COP GSHP
Ltotal
Inv GSHP
COP GSHP
Inv PV
nPV
OCgrid
OCgrid
OCgrid
OCgas
OCgas
OCgas
10 kgCO 2/m 2 5 kgCO 2/m 2
105 105
2 0.6 2 0.6
0.5 0.5
Sobol indices
Sobol indices
0.4 0.4
*
*
1 0.3 1 0.3
0.2 0.2
0.1 0.1
0 0 0 0
nnet
nnet
nGAS
nBIO
Ltotal
Inv GSHP
COP GSHP
Inv PV
OCbio
Ltotal
OCbio
Inv GSHP
COP GSHP
Inv PV
OCgrid
cgrid
OCgrid
cgrid
OCgas
OCgas
Fig. 12. Results of the method of Morris and of the Variance-based Sobol method indicating the most influencing uncertain parameters with regards to system costs for the different
scenarios (Ltotal refers to the energy demand and solar radiation patterns, OC refers to energy carrier prices, Inv refers to the investment cost of an energy technology, n and COP refer to
the efficiencies of an energy technology).
the most important one with OCgas falling at the second and third place
in the two scenarios, respectively. Finally, in Scenarios 4 and 5, OCgas
becomes relatively insignificant, while the price of the biomass energy 0.6
carrier, OCbio, increases in importance. Overall, the energy demands/
solar radiation Ltotal are identified as universally important, while de-
pending on the emission limits, different relative importance is assigned 0.4
to the energy carrier prices OCgrid, OCgas, OCbio.
Besides the most important parameters, GSA results also indicate
that the uncertainty of parameters like investment costs and technical 0.2 Full uncertain parameter set
characteristics can be safely neglected. Hence, these parameters can be Screened uncertain parameter set
fixed at their deterministic values without significant loss of output
0
variance. 20 25 30 35 40 45
As mentioned earlier, the two-step GSA approach reduces the
number of required model runs by almost 50%. To investigate whether Total cost (CHF/m2 )
the screening of the ten most important parameters leads to any loss of Fig. 13. Comparison of Scenario 3 Total cost Cumulative Density Functions (CDF) from
output variability, Fig. 13 compares the Cumulative Distribution Uncertainty Analysis (UA) using the full uncertain parameter set and from Sobol Analysis
Functions of the system cost for Scenario 3 resulting from the MC model using the screened parameter set by the method of Morris.
evaluations of UA and the Sobol method. For the former, all uncertain
parameters are included, while for the latter only the ten most im- parameters and the rest are ignored, no significant loss of accuracy
portant parameters identified by the Morris method. The good agree- would occur.
ment shown between the two curves indicates that indeed the majority
of output uncertainty is accounted for by the ten parameters screened
with the Morris method. 5. Conclusions
Overall, GSA results offer valuable information to DES designers
about the drivers of uncertainty. For the case study examined, for in- This paper presents a four-step framework for the investigation of
stance, results indicate that the DES designer should aim for a reduction uncertainty in optimal DES design. In the first step, a deterministic
of the uncertainty of the energy demands and the energy carrier prices. model is developed for optimal DES design and operation. The second
This would, in return, greatly reduce the uncertainty of the model’s step pertains to UC, in which probabilistic descriptions are assigned to
output i.e. the uncertainty over which design to select as the optimal the model’s uncertain parameters, which are energy carrier prices and
one. Finally, GSA can also inform further modelling efforts. For in- emission factors, investment costs and technical characteristics, and
stance, in the case that a model based on Stochastic Programming (SP) energy demand and solar radiation patterns. In the third step, UA is
for decision-making under uncertainty were to be developed for this performed using Monte Carlo simulations to investigate the variability
paper’s case study, the results would indicate that if only the energy of model outputs. Finally, the fourth step includes a two-stage GSA
carrier prices and the energy demands are considered as uncertain approach combining the Morris and the variance-based Sobol methods
230
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
to determine the most influencing parameters driving the total cost of design outcome. Moreover, the use of these techniques, especially in
the DES. practice, can lead to more robust energy systems against uncertainty.
The framework is applied to an optimal DES design task for a Swiss DES designers can use UA to test the “robustness” of their designs by
urban neighbourhood, in which cost-optimal systems are sought that examining whether they remain optimal for different realisations of the
achieve progressively better emission performance. The deterministic uncertain parameters. Such investigations of uncertainty can also assist
results reveal that while a GSHP-CHP-gas boiler configuration achieves with helping interested parties to invest in DES. This could in turn ac-
the lowest cost, a gradual phase-out of the CHP engine is required to celerate the adoption of DES and contribute to real-world cost, emission
reduce emissions and biomass is only required for CO2 emission per- and energy savings.
formance below 5 kgCO2/m2. UA, though, reveals significant variability Regarding future work, UA and GSA could be used to investigate
in the model outputs and deviations from the deterministic results. uncertainty in different settings, such as for the design of multi-site
More specifically, in many MC simulations, the total system cost is DES. Additionally, models based on techniques for Optimisation under
higher than the deterministic cost estimates. Moreover, the determi- Uncertainty (e.g. Stochastic Programming) can be developed for op-
nistic configurations are shown to be frequently suboptimal choices, timal decision-making under uncertainty. The results of the GSA can be
confirming, thus, the risks of deterministic design. Having established directly fed into those models as they indicate the most influencing
the significant impacts of uncertainty, GSA reveals that the main drivers uncertain parameters that need to be considered.
behind the variability of the optimal systems’ cost are the energy de-
mand patterns and the energy carrier prices with the rest of the un- Acknowledgements
certain parameters having only minimal influence.
Overall, given the multi-faceted nature of uncertainty in DES design, The research in this paper is supported in part by funds from the
maintaining deterministic practices is insufficient. The framework Competence Center Energy and Mobility (CCEM) project IDEAS4cities
presented in this paper, combining a deterministic DES design model (CCEM 801), and the Swiss Commission for Technology and Innovation
with UA and GSA techniques can be adopted by DES researchers and within the Swiss Competence Centres for Energy Research (SCCER)
practitioners to help them understand the range of optimal designs and Future Energy Efficient Buildings & Districts (FEEB&D) (CTI
their performance, and the parameters with the greatest effect on the 1155002539).
Appendix A
In this appendix, the formulation of the deterministic DES design model is provided. The model data and equations are indexed over the sets
presented in Table A.1.
Table A.1
Model sets and indices used for equation formulations.
231
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
where r is the discount rate, and lifetimei corresponds to the lifetime of technology i in years.
The term DH (Eq. (A.3)) represents the district heating network cost composed of the network piping cost and of the cost of the heat exchanger
sub-stations in the buildings. peakb is a parameter for the peak heat load of building b [kW], which dictates the size of its heat exchanger, LCHX is a
linear, capacity-dependent cost parameter for the installation of a heat exchanger HX [CHF/kW], CRFHX is the capital recovery factor (CRF) parameter
for the heat exchanger HX, lnet is a parameter for the length of the piping network [m], and LCnet is the linear cost parameter per m of piping network
[CHF/m].
In Eq. (A.4), Im,t is a continuous nonnegative variable representing the consumption of energy carrier m at time step t [kWh], OCm is a parameter for
the price of energy carrier m [CHF/kWh], Ltexp is a continuous nonnegative variable for electricity exported to the grid at time step t [kWh], and FiT is a
parameter for the feed-in-tariffs of exported electricity [CHF/kWh]. The term wt corresponds to the number of time steps that every typical time step
represents and is calculated by the clustering procedure used to define the set of typical days considered for the design, as explained in Section A.2.
A.1.2. Constraints
In this section, the model constraints are presented defining the feasible region of each decision variable.
Carbon emissions constraint: A constraint setting an upper limit to the system’s CO2 emissions is defined in Eq. (A.6), where Cm is the carbon
emission factor parameter for energy carrier m [gCO2/kWh].
⎧ PGRID,t = Igrid,t ⎫
PBIO,t = Ibio,t , ∀t∈T
⎨ ⎬
P + PGAS,t = Igas,t (A.9)
⎩ CHP,t ⎭
Energy storage balances: The energy storage technologies are assumed to only deal with short-term, daily fluctuations and cannot act as seasonal
storages. To describe the energy balances for the storage devices, two constraints are needed, which are expressed as follows:
Qldis
,t ⩽ Qldis,max ·capl , ∀ l ∈ L,t ∈ T (A.16)
Apart from technology related constraints, other limitations are imposed due to “real world” restrictions. In this model, the capacity of PV panels
232
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
is restricted by the maximum allowable roof surface utilisation, the capacity of storage technologies is constrained due to assumed space limitations,
while the GSHP capacity is also constrained due to limited urban space availability for the required boreholes. The constraints are correspondingly
expressed as follows:
∑ capb,k ⩽ Abroof , ∀ b ∈ B
k∈K (A.17)
⎡ ⎤
RS = ⎢ ∑ (Ibio,t ·wt ) + ∑ (SbPV
,t · capPV ,b · nPV · wt ) ⎥ / TEC ·100%
⎣t∈T t ∈ T ,b ∈ B ⎦ (A.23)
The investment costs for the energy technologies are given in Tables A.2 and A.3 reflecting current prices in the Swiss market. Regarding data for
the energy carrier prices and the electricity feed-in tariffs, it is common practice in deterministic models to use values that correspond to the current
conditions at the time of the design (see e.g. [10,13,48,87]). In this work, values for these parameters are sourced from the Swiss Energy Strategy
Table A.2
Investment cost parameters corresponding to the fixed cost (FCi) and the linear cost (LCi) components of heating and cogeneration technologies [92].
Capacity (kW) FCGAS (CHF) LCGAS (CHF/kW) FCBIO (CHF) LCBIO (CHF/kW) FCGSHP (CHF) LCGSHP (CHF/kW) FCCHP (CHF) LCCHP (CHF/kWth)
Table A.3
Investment cost parameters corresponding to the fixed cost (FCi) and the linear cost (LCi) components of solar, energy storage and district heating technologies [92–94].
2050 (SES2050) [63] to reflect current rates in Switzerland and are shown in Table A4.6 The emission factors used consider direct carbon emissions
from sources within the DES (e.g. due to fuel combustion) as well as indirect emissions associated with the consumption of purchased electricity and
do not reflect the overall life-cycle of the energy carrier and the system that utilises it. The values for these factors are also sourced from SES2050 and
are reported in Table A.4.7
6
Prices for 2016 are not given directly in the SES2050 and their values are linearly interpolated between 2015 and 2020 values.
7
The emission factor for electricity is not directly reported in SES2050 and is calculated by dividing the total carbon emissions due to electricity generation in 2016 by the total
electricity demand of Switzerland for the year 2016. Therefore, a country average emission factor in gCO2 per kWh of electricity consumed in Switzerland can be obtained. The same
approach was also used in [88].
233
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
The values of the technical parameters of the model are given in Tables A.5–A.7. The district heating network length, lnet , is equal to 196 m. The
total building roof area for PV is equal to 1260 m2. The maximum capacity for thermal and electrical storage are taken as 600 kWh (Eq. (A.18)),
while for the GSHP it is taken as equal to 250 kW (Eq. (A.19)). The discount rate r is taken as 8%.
The building energy requirements for heating and electricity and the incoming solar radiation patterns are calculated using the Building
Performance Simulation software EnergyPlus [73]. Building construction and infiltration information are taken from [76,77]. Information regarding
occupancy densities, capacities for lighting and electrical equipment, thermostat and ventilation settings, and hot water demands, and their cor-
responding schedules are taken from the Swiss norm SIA 2024 [78]. Finally, as weather conditions a Typical Meteorological Year (TMY) from [89] is
used. Having information about energy demands and solar radiation for a full calendar year, a set of typical days is generated using the clustering-
based (k-medoids) method in [90]. In total, 12 + 2 days are selected to form the typical set following literature recommendations [90,91]. The
12 days are representative days of the whole year and the 2 additional days correspond to the days when the peak heat and electricity demands are
observed and are maintained for accurate system sizing.
Table A.4
Energy prices and emission factors for the different energy carriers [63].
1
The Rp. is the hundredth denomination of the Swiss franc currency (CHF).
Table A.5
Conversion efficiency, lifetime and minimum available capacities of energy technologies.
Conversion technologies
Conversion efficiency < 50 kW nGAS = 85% nBIO = 80% COPGSHP = 3.5 nCHP,th/ nCHP,el = 52%/33% nPV = 15%
50–200 kW nGAS = 87.5% nBIO = 82.5% COPGSHP = 3.5 nCHP,th/ nCHP,el = 53.5%/34%
> 200 kW nGAS = 87.5% nBIO = 85% COPGSHP = 3.5 nCHP,th/ nCHP,el = 55%/35%
Table A.6
Technical characteristics and lifetime of energy storage modules [95].
Table A.7
Efficiency and lifetime of district heating network components [94,96].
234
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Appendix B
Table B.1
Summary of uncertain parameters and their characterisation approaches for this paper (The parameter μ for the normal (N) and the half-normal (HN) distributions refer to the nominal,
deterministic value).
Appendix C
In this Appendix, the technical details of the GSA method of Morris and the Variance-based Sobol method are given. Extensive information on
these methods can be found in [35].
The method of Morris [51] includes the Monte Carlo evaluation of a model for points along a multi-step trajectory in the input parameter space.
This trajectory is created by modifying one randomly selected input parameter per step by an amount Δ, while keeping the rest constant. The process
is repeated until all k uncertain parameters have moved exactly one step each, bringing the total number of trajectory points to k + 1. As in each step
of the trajectory only a single parameter value xi is altered, its Elementary Effect (EE) can be calculated as follows:
g (x1,⋯,x i − 1,x i ± Δ,⋯,xk )−g (x1,⋯,xk )
EEi =
Δ (C.1)
By defining r different trajectories and performing the same procedure, r elementary effects can be calculated per parameter. The computational
cost for the complete procedure then becomes equal to r(k + 1) model evaluations with a suggested value for r being 4–10 [35].
The original method by Morris then includes the calculation of the mean, μi and the standard deviation, σi, of the EEi as measures of parameter
importance. However, in cases when the elementary effects of one parameter have opposite signs and cancel each other out, small μi values might
emerge, which would falsely indicate a non-influential parameter. As a result, Campolongo et al. [97] have suggested a modified mean measure μi∗,
which considers the absolute values of the elementary effects:
r
1
μi∗ = ∑ |EEi |
r j=1 (C.2)
235
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Campolongo et al. [97] have demonstrated that the measure μi∗ can be used very effectively to rank parameters. However, it is a qualitative
measure and cannot indicate how much more important one parameter is over another. Nevertheless, calculating the measure μi∗ for each parameter
allows for uninfluential parameters to be identified and removed in order for the most computationally intensive Variance-based Sobol method to be
performed.
All the calculations for the Morris screening have been performed using the morris() function from the R package sensitivity [98] that
incorporates all of the method aspects discussed.
The variance-based Sobol method is based on variance decomposition and its aim is the quantification of the contribution of each parameter to
the total variance of the output. For that, the most common practice involves the calculation of two different metrics named first-order Sobol index, Si ,
and total-order Sobol index, STi , respectively:
VXi (E X∼ i (Y |Xi ))
Si =
V (Y ) (C.3)
E X∼ i (VXi (Y |X∼ i ))
STi =
V (Y ) (C.4)
In Eq. (C.3), the term E X∼ i (Y |Xi ) takes the mean of Y overall all the possible values of X∼ i when parameter Xi is fixed. The term X∼ i represents all
input parameters but Xi . The outer variance term VXi is then considered over all values of Xi , thus, giving the variance of the inner mean term as Xi
varies. A more intuitive interpretation of the Si index is that it represents the “expected reduction in the output variance that would be obtained if Xi could
be fixed” [34]. As a result, this index is used to identify the uncertain parameters that have significant contribution to the output’s variance. However,
the first-order indices do not consider any interactions between parameters and only quantify the effect of a parameter by itself. Therefore, a small
first-order index Si cannot indicate non-significance for the parameter Xi as there might exist higher-order interactions.
The total-order Sobol index aims to quantify the total effect of a parameter to the output considering interactions with other parameters. In Eq.
(C.4), the term VXi (Y |X∼ i ) is the variance of the output Y over all possible values for Xi when X∼ i is fixed. The outer term E X∼ i considers all the
possible values of X∼ i , hence, giving the expected variance of the output when X∼ i is fixed. Similarly to the case of the first-order index, Saltelli and
Annoni [34] offer a more intuitive explanation writing that the term E X∼ i (VXi (Y |X∼ i )) “would represent the expected variance that would remain if all
factors but Xi could be fixed”. If for a parameter Xi , STi is equal to zero, this would mean that no variance in the output would remain; hence, despite
the fact that Xi remains uncertain, it has no influence on the output. As a result, the condition STi = 0 is necessary and sufficient for a parameter to be
deemed non-influential.
By definition, the sum of the first-order indices cannot exceed unity and it represents the percentage of the output variance that can be explained
due to inputs taken independently. The difference between the sum of the first-order indices and unity indicates the contribution of interactions
between the parameters.
In order to estimate the first- and total-order Sobol indices, methods based on MC simulations have been proposed [99-101]. All calculations in
this paper are performed using the R package sensitivity [98] that implements all estimation formulae in the functions sobol2007() and so-
boljansen().
Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.apenergy.2018.01.062.
236
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
Res Des 2013;91:2054–69. http://dx.doi.org/10.1016/j.cherd.2013.08.016. [47] Yoshida S, Ito K, Yokoyama R. Sensitivity analysis in structure optimization of
[19] Allegrini J, Orehounig K, Mavromatidis G, Ruesch F, Dorer V, Evins R. A review of energy supply systems for a hospital. Energy Convers Manag 2007;48:2836–43.
modelling approaches and tools for the simulation of district-scale energy systems. http://dx.doi.org/10.1016/j.enconman.2007.06.045.
Renew Sustain Energy Rev 2015;52:1391–404. http://dx.doi.org/10.1016/j.rser. [48] Zhou Z, Liu P, Li Z, Ni W. Economic assessment of a distributed energy system in a
2015.07.123. new residential area with existing grid coverage in China. Comput Chem Eng
[20] Carvalho M, Lozano MA, Serra LM. Multicriteria synthesis of trigeneration systems 2013;48:165–74. http://dx.doi.org/10.1016/j.compchemeng.2012.08.013.
considering economic and environmental aspects. Appl Energy 2012;91:245–54. [49] Li L, Mu H, Gao W, Li M. Optimization and analysis of CCHP system based on
http://dx.doi.org/10.1016/j.apenergy.2011.09.029. energy loads coupling of residential and office buildings. Appl Energy
[21] Marquant JF, Evins R, Bollinger LA, Carmeliet J. A holarchic approach for multi- 2014;136:206–16. http://dx.doi.org/10.1016/j.apenergy.2014.09.020.
scale distributed energy system optimisation. Appl Energy 2017;208:935–53. [50] Moret S, Bierlaire M, Maréchal F. Robust optimization for strategic energy plan-
http://dx.doi.org/10.1016/j.apenergy.2017.09.057. ning. Lausanne: Transport and Mobility Laboratory; 2014.
[22] Gabrielli P, Gazzani M, Martelli E, Mazzotti M. Optimal design of multi-energy [51] Morris MD. Factorial sampling plans for preliminary computational experiments.
systems with seasonal storage. Appl Energy 2017. http://dx.doi.org/10.1016/j. Technometrics 1991;33:161–74. http://dx.doi.org/10.2307/1269043.
apenergy.2017.07.142. [52] Sudret B. Uncertainty propagation and sensitivity analysis in mechanical models.
[23] Wu R, Mavromatidis G, Orehounig K, Carmeliet J. Multiobjective optimisation of Contributions to structural reliability and stochastic spectral methods. Clermont-
energy systems and building envelope retrofit in a residential community. Appl Ferrand: Université Blaise-Pascal; 2007.
Energy 2017;190:634–49. http://dx.doi.org/10.1016/j.apenergy.2016.12.161. [53] Geidl M, Koeppel G, Favre-Perrod P, Klockl B, Andersson G, Frohlich K. Energy
[24] Stadler M, Groissböck M, Cardoso G, Marnay C. Optimizing Distributed Energy hubs for the future. IEEE Power Energy Mag 2007;5:24–30. http://dx.doi.org/10.
Resources and building retrofits with the strategic DER-CAModel. Appl Energy 1109/MPAE.2007.264850.
2014;132:557–67. http://dx.doi.org/10.1016/j.apenergy.2014.07.041. [54] Evins R, Orehounig K, Dorer V, Carmeliet J. New formulations of the ‘energy hub’
[25] Ren H, Zhou W, Gao W. Optimal option of distributed energy systems for building model to address operational constraints. Energy 2014;73:387–98. http://dx.doi.
complexes in different climate zones in China. Appl Energy 2012;91:156–65. org/10.1016/j.energy.2014.06.029.
http://dx.doi.org/10.1016/j.apenergy.2011.08.044. [55] Orehounig K, Evins R, Dorer V. Integration of decentralized energy systems in
[26] Di Somma M, Yan B, Bianco N, Graditi G, Luh PB, Mongibello L, et al. Multi- neighbourhoods using the energy hub approach. Appl Energy 2015;154:277–89.
objective design optimization of distributed energy systems through cost and ex- http://dx.doi.org/10.1016/j.apenergy.2015.04.114.
ergy assessments. Appl Energy 2017;204:1299–316. http://dx.doi.org/10.1016/j. [56] Mavromatidis G, Orehounig K, Carmeliet J. Evaluation of photovoltaic integration
apenergy.2017.03.105. potential in a village. Sol Energy 2015;121:152–68. http://dx.doi.org/10.1016/j.
[27] Kienzle F, Ahcin P, Andersson G. Valuing investments in multi-energy conversion, solener.2015.03.044.
storage, and demand-side management systems under uncertainty. IEEE Trans [57] Mavromatidis G, Evins R, Orehounig K, Dorer V, Carmeliet J. Multi-objective
Sustain Energy 2011;2:194–202. http://dx.doi.org/10.1109/TSTE.2011.2106228. optimization to simultaneously address energy hub layout, sizing and scheduling
[28] Mavrotas G, Florios K, Vlachou D. Energy planning of a hospital using using a linear formulation. In: Araujo A, editor. Eng. Optim. 2014, CRC Press;
Mathematical Programming and Monte Carlo simulation for dealing with un- 2014, p. 603–8.
certainty in the economic parameters. Energy Convers Manag 2010;51:722–31. [58] Haimes Y, Ladson LS, Wismer DA. On a bicriterion formulation of the problems of
http://dx.doi.org/10.1016/j.enconman.2009.10.029. integrated system identification and system optimization. IEEE Trans Syst Man
[29] Lythcke-Jørgensen C, Ensinas AV, Münster M, Haglind F. A methodology for de- Cybern 1971;1:296–7. http://dx.doi.org/10.1109/TSMC.1971.4308298.
signing flexible multi-generation systems. Energy 2016;110:34–54. http://dx.doi. [59] Aven T. Interpretations of alternative uncertainty representations in a reliability
org/10.1016/j.energy.2016.01.084. and risk analysis context. Reliab Eng Syst Saf 2011;96:353–60. http://dx.doi.org/
[30] Hawkes AD, Leach MA. Modelling high level system design and unit commitment 10.1016/j.ress.2010.11.004.
for a microgrid. Appl Energy 2009;86:1253–65. http://dx.doi.org/10.1016/j. [60] Chowdhury S, Champagne P, McLellan PJ. Uncertainty characterization ap-
apenergy.2008.09.006. proaches for risk assessment of DBPs in drinking water: a review. J Environ
[31] Li C-Z, Shi Y-M, Liu S, Zheng Z, Liu Y. Uncertain programming of building cooling Manage 2009;90:1680–91. http://dx.doi.org/10.1016/j.jenvman.2008.12.014.
heating and power (BCHP) system based on Monte-Carlo method. Energy Build [61] Mavromatidis G, Orehounig K, Carmeliet J. A review of uncertainty character-
2010;42:1369–75. http://dx.doi.org/10.1016/j.enbuild.2010.03.005. isation approaches for the optimal design of distributed energy systems. Renew
[32] Dufo-López R, Pérez-Cebollada E, Bernal-Agustín JL, Martínez-Ruiz I. Optimisation Sustain Energy Rev, submitted for publication.
of energy supply at off-grid healthcare facilities using Monte Carlo simulation. [62] Mirkhani S, Saboohi Y. Stochastic modeling of the energy supply system with
Energy Convers Manag 2016;113:321–30. http://dx.doi.org/10.1016/j.enconman. uncertain fuel price – a case of emerging technologies for distributed power gen-
2016.01.057. eration. Appl Energy 2012;93:668–74. http://dx.doi.org/10.1016/j.apenergy.
[33] Saltelli A, editor. Sensitivity analysis in practice: a guide to assessing scientific 2011.12.099.
models. Hoboken, NJ: Wiley; 2004. [63] Prognos AG. Die Energieperspektiven fur die Schweiz bis 2050. Swiss Federal
[34] Saltelli A, Annoni P. How to avoid a perfunctory sensitivity analysis. Environ Office of Energy SFOE (Bundesamt für Energie BFE); 2012.
Model Softw 2010;25:1508–17. http://dx.doi.org/10.1016/j.envsoft.2010.04.012. [64] Densing M, Hirschberg S, Turton H. Review of Swiss electricity scenarios 2050.
[35] Saltelli A, editor. Global sensitivity analysis: the primer. Chichester, England; PSI–Report 14-05 Paul Scherrer Institute, Villigen PSI, Switzerland (2014) ISSN
Hoboken, NJ: John Wiley; 2008. 1019-0643; 2014. Available from: < http://www.psi.ch/eem/
[36] Borgonovo E, Plischke E. Sensitivity analysis: a review of recent advances. Eur J PublicationsTabelle/PSI-Bericht_14-05.pdf > .
Oper Res 2016;248:869–87. http://dx.doi.org/10.1016/j.ejor.2015.06.032. [65] SFOE. Schweizerische Gesamtenergiestatistik 2015. Swiss Federal Office of Energy
[37] Iooss B, Lemaître P. A review on global sensitivity analysis methods. In: Dellino G, (SFOE); 2015. Available from: < http://www.bfe.admin.ch/themen/00526/
Meloni C, editors. Uncertain. Manag. Simul.-Optim. Complex Syst., vol. 59, Boston, 00541/00542/00631/index.html?lang=en&dossier_id=00763 > .
MA: Springer US; 2015, p. 101–22. [66] Carpentiero V, Langella R, Testa A. Hybrid wind-diesel stand-alone system sizing
[38] Tian W. A review of sensitivity analysis methods in building energy analysis. accounting for component expected life and fuel price uncertainty. Electr Power
Renew Sustain Energy Rev 2013;20:411–9. http://dx.doi.org/10.1016/j.rser. Syst Res 2012;88:69–77. http://dx.doi.org/10.1016/j.epsr.2012.02.003.
2012.12.014. [67] IEA. Energy Policies of IEA Countries – Switzerland. International Energy Agency
[39] Ashouri A, Petrini F, Bornatico R, Benz MJ. Sensitivity analysis for robust design of (IEA), Paris, France; 2012.
building energy systems. Energy 2014;76:264–75. http://dx.doi.org/10.1016/j. [68] Burhenne S, Tsvetkova O, Jacob D, Henze GP, Wagner A. Uncertainty quantifi-
energy.2014.07.095. cation for combined building performance and cost-benefit analyses. Build Environ
[40] Bracco S, Dentici G, Siri S. DESOD: a mathematical programming tool to optimally 2013;62:143–54. http://dx.doi.org/10.1016/j.buildenv.2013.01.013.
design a distributed energy system. Energy 2016;100:298–309. http://dx.doi.org/ [69] Giannakoudis G, Papadopoulos AI, Seferlis P, Voutetakis S. Optimum design and
10.1016/j.energy.2016.01.050. operation under uncertainty of power systems using renewable energy sources and
[41] Bustos F, Toledo A, Contreras J, Fuentes A. Sensitivity analysis of a photovoltaic hydrogen storage. Int J Hydrog Energy 2010;35:872–91. http://dx.doi.org/10.
solar plant in Chile. Renew Energy 2016;87(Part 1):145–53. http://dx.doi.org/10. 1016/j.ijhydene.2009.11.044.
1016/j.renene.2015.09.070. [70] de Wit S. Uncertainty in building simulation. In: Augenbroe G, Malkawi A, editors.
[42] Lindberg KB, Doorman G, Fischer D, Korpås M, Ånestad A, Sartori I. Methodology Adv. Build. Simul., Oxford: Taylor & Francis; 2003.
for optimal energy system design of Zero Energy Buildings using mixed-integer [71] Gang W, Augenbroe G, Wang S, Fan C, Xiao F. An uncertainty-based design op-
linear programming. Energy Build 2016;127:194–205. http://dx.doi.org/10. timization method for district cooling systems. Energy 2016;102:516–27. http://
1016/j.enbuild.2016.05.039. dx.doi.org/10.1016/j.energy.2016.02.107.
[43] Mehleri ED, Sarimveis H, Markatos NC, Papageorgiou LG. Optimal design and [72] Hopfe CJ, Hensen JLM. Uncertainty analysis in building performance simulation
operation of distributed energy systems: application to Greek residential sector. for design support. Energy Build 2011;43:2798–805. http://dx.doi.org/10.1016/j.
Renew Energy 2013;51:331–42. http://dx.doi.org/10.1016/j.renene.2012.09.009. enbuild.2011.06.034.
[44] Ren H, Gao W, Ruan Y. Optimal sizing for residential CHP system. Appl Therm Eng [73] EnergyPlus. EnergyPlus Simulation Software; 2016. Available from: < https://
2008;28:514–23. http://dx.doi.org/10.1016/j.applthermaleng.2007.05.001. energyplus.net/ > .
[45] Ren H, Gao W. A MILP model for integrated plan and evaluation of distributed [74] Macdonald IA. Quantifying the effects of uncertainty in building simulation.
energy systems. Appl Energy 2010;87:1001–14. http://dx.doi.org/10.1016/j. University of Strathclyde; 2002.
apenergy.2009.09.023. [75] Ioannou A, Itard LCM. Energy performance and comfort in residential buildings:
[46] Rubio-Maya C, Uche-Marcuello J, Martínez-Gracia A, Bayod-Rújula AA. Design sensitivity for building parameters and occupancy. Energy Build 2015;92:216–33.
optimization of a polygeneration plant fuelled by natural gas and renewable en- http://dx.doi.org/10.1016/j.enbuild.2015.01.055.
ergy sources. Appl Energy 2011;88:449–57. http://dx.doi.org/10.1016/j. [76] Landolt J. A bottom-up modelling approach to address sustainable transformation
apenergy.2010.07.009. strategies for the Swiss building stock. ETH Zurich; 2016.
237
G. Mavromatidis et al. Applied Energy 214 (2018) 219–238
[77] Wang D, Landolt J, Orehounig K, Carmeliet J. Dynamic urban energy demand Selection of typical demand days for CHP optimization. Energy Build
modelling to address building retrofitting alternatives in Switzerland. Status- 2011;43:3036–43. http://dx.doi.org/10.1016/j.enbuild.2011.07.024.
Semin. 2016 Zurich Switz.; 2016. [91] Schütz T, Schraven H, Harb H, Fuchs M, Müller D. Clustering algorithms for the
[78] SIA. SIA 2024 Standard-Nutzungsbedingungen fuer die Energie- und selection of typical demand days for the optimal design of building energy systems.
Gebaeudetechnik. Swiss society of Engineers and Architects; 2006. Proc. ECOS 2016; 2016.
[79] Gang W, Wang S, Shan K, Gao D. Impacts of cooling load calculation uncertainties [92] Jakob M, Kallio S, Nägeli, C, Ott W, Bolliger, R, von Grünigen S. INSPIRE-Tool.
on the design optimization of building cooling systems. Energy Build 2015;94:1–9. TEP Energy GmbH & econcept AG. Supported by the Swiss Federal Office of Energy
http://dx.doi.org/10.1016/j.enbuild.2015.02.032. (SFOE) – EnergieSchweiz; 2015. Available from: < http://www.inspire-tool.ch > .
[80] IPCC. Climate Change 2014: Synthesis Report. Contribution of Working Groups I, [93] Jenni Energietechnik. Hot or Cold Storage Systems; 2016. Available from: <
II and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate http://www.jenni.ch/files/jenni/inhalte/pdf/Produkte/Preise_Solartank.pdf > .
Change [Core Writing Team, R.K. Pachauri and L.A. Meyer (eds.)]. IPCC, Geneva, [94] Zernez. Zernez Energia 2020; 2015. Available from: < http://zernezenergia2020.
Switzerland; 2014. 151 pp. ch/2015 > .
[81] Giorgi F, Jones C, Asrar GR. Addressing climate information needs at the regional [95] Marnay C, Venkataramanan G, Stadler M, Siddiqui AS, Firestone R, Chandran B.
level: the CORDEX framework. World Meteorol Organ WMO Bull 2009;58:175. Optimal technology selection and operation of commercial-building microgrids.
[82] Belcher S, Hacker J, Powell D. Constructing design weather data for future cli- IEEE Trans Power Syst 2008;23:975–82. http://dx.doi.org/10.1109/TPWRS.2008.
mates. Build Serv Eng Res Technol 2005;26:49–61. http://dx.doi.org/10.1191/ 922654.
0143624405bt112oa. [96] Gudmundsson O, Thorsen JE, Zhang L. Cost analysis of district heating compared
[83] Mathworks. MATLAB 2016b. The MathWorks, Inc., Natick, Massachusetts, USA; to its competing technologies; 2013, p. 3–13. 10.2495/ESUS130091.
2016. [97] Campolongo F, Cariboni J, Saltelli A. An effective screening design for sensitivity
[84] Kucherenko S, Albrecht D, Saltelli A. Exploring multi-dimensional spaces: a analysis of large models. Environ Model Softw 2007;22:1509–18. http://dx.doi.
Comparison of Latin Hypercube and Quasi Monte Carlo Sampling Techniques. org/10.1016/j.envsoft.2006.10.004.
ArXiv150502350 Stat; 2015. [98] Pujol G, Iooss B, Boumhaout AJ with contributions from K, Veiga SD, Fruth J,
[85] Homma T, Saltelli A. Importance measures in global sensitivity analysis of non- Gilquin L, et al. sensitivity: Global Sensitivity Analysis of Model Outputs. R
linear models. Reliab Eng Syst Saf 1996;52:1–17. http://dx.doi.org/10.1016/ package version 1.12; 2016. Available from: < http://CRAN.R-project.org/
0951-8320(96)00002-6. package=sensitivity > .
[86] Williams HP. Model building in mathematical programming. 5th ed. Hoboken, [99] Saltelli A, Annoni P, Azzini I, Campolongo F, Ratto M, Tarantola S. Variance based
N.J: Wiley; 2013. sensitivity analysis of model output. Design and estimator for the total sensitivity
[87] Morvaj B, Evins R, Carmeliet J. Optimization framework for distributed energy index. Comput Phys Commun 2010;181:259–70. http://dx.doi.org/10.1016/j.cpc.
systems with integrated electrical grid constraints. Appl Energy 2009.09.018.
2016;171:296–313. http://dx.doi.org/10.1016/j.apenergy.2016.03.090. [100] Sobol’ IM, Tarantola S, Gatelli D, Kucherenko SS, Mauntz W. Estimating the ap-
[88] Mavromatidis G, Orehounig K, Richner P, Carmeliet J. A strategy for reducing CO2 proximation error when fixing unessential factors in global sensitivity analysis.
emissions from buildings with the Kaya identity – a Swiss energy system analysis Reliab Eng Syst Saf 2007;92:957–60. http://dx.doi.org/10.1016/j.ress.2006.07.
and a case study. Energy Policy 2016;88:343–54. http://dx.doi.org/10.1016/j. 001.
enpol.2015.10.037. [101] Jansen MJW. Analysis of variance designs for model output. Comput Phys
[89] Meteotest. Meteonorm. Bern, Switzerland: Meteotest; 2015. Commun 1999;117:35–43. http://dx.doi.org/10.1016/S0010-4655(98)00154-4.
[90] Domínguez-Muñoz F, Cejudo-López JM, Carrillo-Andrés A, Gallardo-Salazar M.
238