Professional Documents
Culture Documents
Abstract—In the ever complex world of the power systems, The use of flexibility in form of demand flexibility, storage,
robust decision making in the context of investment planning or power flow control through High Voltage Direct Current
becomes a difficult problem. In the decision making process, (HVDC) links can help to decrease the need for classical grid
the costs and benefits of implementing complementary or com-
peting technologies (AC versus DC or storage versus classical expansion and achieve significant cost savings considering the
transmission expansion) need to be assessed. Additionally, the large investment volumes projected. Considering the difficul-
electrification of industrial demand and the integration of smart ties for obtaining permissions for large infrastructure projects,
loads, such as heat pumps or electric vehicles, pave the way for making use of flexibility becomes even more important.
demand flexibility provision on large scale. Considering that the
majority of the demand flexibility is located in the distribution
grid, there is a need for planning models and tools able to In order to find the necessary trade-offs between classical
perform combined transmission and distribution grid planning. grid investment – into both transmission and distribution net-
This paper introduces FlexPlan.jl, a Julia/JuMP-based open- works – and sources of flexibility, a holistic planning model is
source tool for holistic planning of transmission and distribution needed. Although many transmission expansion optimisation
2023 Open Source Modelling and Simulation of Energy Systems (OSMSES) | 979-8-3503-1121-1/23/$31.00 ©2023 IEEE
grids which includes a complete set of planning candidates for models exist in the literature, to the best of our knowledge
transmission and distribution networks and fully internalizes
demand flexibility and storage usage. We use stochastic op- there are no tractable implementations of the problem, which
timisation, in order to find robust decisions with respect to can consider all layers of the network, and a large uncertainty
different climate conditions and operating hours within given set to represent future operation conditions with respect to
climate years. To keep the optimisation problem tractable, we renewable generation and demand.
introduce a novel decomposition between the transmission and
distribution grid planning problems. We demonstrate that using
the proposed approach a speed improvement of up to 100 times The aim of the FlexPlan project is to tackle this issue, by
can be achieved for cases with a large number of distribution providing a comprehensive planning model [4]. Within the
grids, with negligible increase of the objective function value and project, a chain of tools have been developed for creating
a good solution quality in terms of power flow. realistic planning scenarios, for determining the location and
Index Terms—Holistic planning, mixed-integer optimisation, size of expansion candidates, and to perform the planning op-
transmission grid planning, distribution grid planning, demand
flexibility, storage, hvdc. timisation. Within the tool-chain, FlexPlan.jl builds the rapid
prototyping platform that has been developed to test different
problem formulations, equipment and flexibility models, and
I. I NTRODUCTION
solution approaches [5]. At the same time FlexPlan.jl provides
The need for decarbonising the electricity system and the a reference implementation to the cloud-based planning tool
inherent increase in renewable generation requires massive developed within FlexPlan [6]. FlexPlan.jl has been released
investments into electricity transmission and distribution grids as a registered open-source Julia package under BSD 3-Clause
in the near future. According to the European Commission, License.
800 GC worth of investments are needed for boosting the
offshore renewable generation by 2050, of which two thirds The next section describes the design specifications of
are expected to be grid infrastructure investments [1]. Sim- FlexPlan.jl as a rapid prototyping platform for grid planning.
ilarly, according to Eurelectric, distribution grid investments In Section III, the scope of the planning model is described,
of 375 ÷ 475 GC are required, in order to facilitate the outlining the model structure, component and network mod-
utilisation of distributed generation, electrification of transport els. Further, in Section IV, a novel approach to decompose
and domestic heat demand [2]. More importantly, ENTSO- transmission and distribution grid planning models is shown,
E states that annual network investments of 1.3 GC could which provides accurate results with significant improvement
facilitate the decrease of power generation costs by 4 GC per of computation time with respect to the combined model.
year [3], demonstrating that grid expansion is inevitable for Finally, Section V draws the main conclusions and provides
affordable energy supply in the future. directions for future work.
III. D ESCRIPTION OF THE GENERIC PLANNING MODEL Here, T is the set of operation periods, whereas Y is the
The outline of the optimisation model implemented in set of investment periods (“planning years”). The stochastic
FlexPlan.jl is provided in Fig. 1. In general, the optimisation dimension is introduced by weighting the operation cost of
model uses three sets of inputs. Firstly, the grid data for the each possible scenario s ∈ S with its probability πs , such that
existing transmission and distribution networks are required. the optimization process returns a unique set of investment
The data models of [9] and [12] are used to represent AC decisions with the statistically highest benefits in terms of total
and DC grids respectively. Secondly, a list of candidate grid costs containment.
expansion options need to be provided, namely candidate AC Investment costs are computed as
and DC grid extensions, demand flexibility and storage invest- X
cc cc
X
dc dc
ments from which the optimal subset is determined. Lastly, [CAPEX]y = Cc,y αc,y + Cd,y αd,y +
c∈S cc d∈S dc
as planning scenarios, a number of renewable generation and X X
ac ac bc bc
demand time series are required, defined per generator and per + Ca,y αa,y + Cb,y αb,y + (2)
load. The length of the time series can be determined by the a∈S ac b∈S bc
user. For sequential planning, the time series can be defined for
X lf lf
+ Cl,y αl,y
multiple planning years. Further, multiple sets of time series l∈S lf
∗ ∗
where C·,y represent cost parameters and α·,y are binary parameters, whereas the voltage taps of on-load tap changers
investment decision variables. The model accounts for in- in distribution networks are modelled as continuous variables.
vestments in grid expansion elements: DC converters (S cc ); 3) Generators modelling: The power produced by dis-
DC branches (S dc ); AC branches (S ac ) such as lines, cables, patchable generators g ∈ S gd is bounded by a parameter:
g g,max
and transformers; storage devices (S bc ); as well as for the Pg,t,s,y ∈ [0, Pg,t,s,y ]. Non-dispatchable generators g ∈ S gn ,
flexibilization (S lf ) of existing loads. gref
instead, have an input reference power Pg,t,s,y and can be
Operation costs are defined by curtailed, so their produced power is
X g g
[OPEX]t,s,y = Cg,t,s,y Pg,t,s,y + g
Pg,t,s,y gref
= Pg,t,s,y gcurt
− Pg,t,s,y (4)
g∈S gd
gcurt gref
X gcurt gcurt
X
lcurt lcurt where the curtailed power is bounded: Pg,t,s,y ∈ [0, Pg,t,s,y ].
+ Cg,t,s,y Pg,t,s,y + Cl,t,s,y Pl,t,s,y +
g∈S gn
(3) b
4) Storage modelling: Existing (b ∈ S ) and candidate (b ∈
l∈S l ∪S lf
S bc ) storage is modelled as a realistic lossy storage device. The
!
lsh
X
lred lred
Cl,t,s,y
lshup lshdn
+ Cl,t,s,y Pl,t,s,y + Pl,t,s,y +Pl,t,s,y state equation
2
l∈S lf
∆t
∗
where C·,t,s,y ∗
are cost parameters and P·,t,s,y are continuous Eb,t,s,y = (1 − λb ) Eb,t−1,s,y +
inj
!
decision variables representing per-unit power. Dispatchable Pb,t,s,y (5)
generators g ∈ S gd are accounted for production (g), whereas + ∆t ηbabs Pb,t,s,y
abs
− + ξb,t,s,y
ηbinj
non-dispatchable generators g ∈ S gn can be curtailed (gcurt).
Loads l ∈ S lf can provide flexibility services – namely, computes the energy Eb,t,s,y ∈ [0, Ebmax ] at end of operation
up- (lshup) and downward (lshdn) demand shifting (lsh), period t from the energy at the previous period depending
and voluntary reduction (lred) – upon investment decision on: the self-discharge rate λb ∈ (0, 1); the absorbed/injected
lf abs,max inj inj,max
αl,y . The set S l , instead, represents loads that cannot provide abs
power Pb,t,s,y ∈ [0, Pb,t,s,y ] and Pb,t,s,y ∈ [0, Pb,t,s,y ],
flexibility services. Both types of load can be curtailed (lcurt). which are optimization variables; the absorption/injection ef-
In order to consider the environmental impact, the carbon ficiencies ηbabs ∈ (0, 1) and ηbinj ∈ (0, 1); and the external
footprint costs and the landscape impact can be included in process ξb,t,s,y , an input parameter representing changes in
the investment cost, whereas the CO2 emission costs can stored energy (e.g. rain flow and spillage for hydro reser-
be included in the power generation cost. Details for the voirs). Energy levels Ebinit and Ebf inal at the boundaries of
calculation of the environmental costs can be found under [13]. the optimization horizon are input parameters. Regardless of
To quantify landscape impact related costs, [14] provides an technology (synchronous machine or inverter-based), storage
open-source tool based on the algorithm described in [15]. devices are assumed to be capable of providing reactive
power support. Although only active power is modelled in
B. Component and network models
the transmission system, the modelling of reactive power is
For the sake of brevity, in the following paragraphs we are essential in distribution grid planning problems as outlined in
only briefly describing the component and network models. Section III-B6.
References for the modelling details are provided at the 5) Demand flexibility modelling: The non-negative demand
appropriate locations. of loads, used in the nodal power balance equations as outlined
1) HVDC lines, branches and converters: The HVDC in Section III-B6, is computed as
substation and grid models are based on [12]. The substations
and DC branches are modelled separately in order to allow lf lex lref lshup lshdn lred lcurt
Pl,t,s,y = Pl,t,s,y + Pl,t,s,y − Pl,t,s,y − Pl,t,s,y − Pl,t,s,y (6)
the modelling of meshed DC grids. The substation model
includes converter transformers, reactive and capacitive filters, for loads l ∈ S lf , and as
and a lossy power electronic converter model. DC branches lf lex lref lcurt
are defined to connect a pair of DC nodes. Although different Pl,t,s,y = Pl,t,s,y − Pl,t,s,y (7)
power flow models have been proposed in [12], to keep the lref
model linear, a network flow model is used for DC branches. for loads l ∈ S l . The reference demand parameter Pl,t,s,y ≥0
As for the DC converter substations, the linearised “DC power gives the base profile, which is time- and scenario-dependent.
flow” formulation is chosen. One of features of the model is The flexibility variables introduced in (3) are all bounded by
the possibility to define candidate DC nodes for modelling DC parameters that can be adapted by the user depending on
lshup lshup,max lshdn
connections between DC nodes that might not yet exist, and the load characteristics: Pl,t,s,y ∈ [0, Pl,t,s,y ], Pl,t,s,y ∈
lshdn,max lred lred,max
also to allow intermediate tappings within DC interconnectors, [0, Pl,t,s,y ], and Pl,t,s,y ∈ [0, Pl,t,s,y ].
which is useful for modelling hybrid offshore assets. It is assumed that demand shifting is re-balanced every T r
2) AC lines, cables and transformers: A generic π-section periods, that is, for every τ ∈ T such that mod (τ, T r ) = 0:
model is used to represent AC branches in general as provided τ
in [9]. For AC lines, cables and power transformers, the model
X lshup lshdn
Pl,t,s,y − Pl,t,s,y = 0. (8)
allows the inclusion of phase angle and voltage taps as fixed t=τ −T r +1
6) Network power flow formulations: To represent meshed Real grid Combined model
Transmission network
AC/DC transmission grids, the well known linearised “DC
power flow” approach is used. Note that due to the choice
(meshed)
of the linearised power flow model, the DC grid power
flows transpose to a network flow model as stated in [12].
Although this introduces some inaccuracy, considering that T PCC bus
meshed HVDC networks are not that wide-spread, the effect T coupling generator
on the overall network will be rather limited. Separate binary PCC bus coupling constraints
investment decision variables have been used for AC branches,
Distribution network
D coupling generator
DC branches and DC converters, allowing to define a large D PCC bus
variety of expansion candidates and thus different grid topolo-
(radial)
gies. Concerning distribution networks, active power exchange
of local demand and generation has a significant impact on the
voltage because of the higher R/X ratio of lines and cables.
This means that, in addition to power transport capacity,
voltage limits can likely represent a bottleneck against load and
Fig. 2. Coupling of the transmission network and a distribution network using
generation increase. As such, the application of the “DC power the combined model.
flow” approach would not allow to investigate the effects of
voltage congestion. To that end, we have chosen to use the sim-
plified DistFlow model [16] to represent distribution networks. A. Combined model
Under the assumptions that the grid is operated radially and In the combined model, coupling is implemented by split-
active/reactive power losses are reasonably small, the selected ting the network at the Points of Common Coupling (PCCs) of
linear formulation is demonstrated to be a good approximation transmission and distribution as shown in Fig. 2. Firstly, PCC
of the exact power flow formulation. Furthermore, in order buses are duplicated, each copy retaining only connections
to consider power transfer capacity of distribution lines and to either transmission or distribution network components.
transformers without neglecting the impact of reactive power Then, a coupling generator, having zero cost and a rectangular
flows, the apparent-power circular-constraint is linearized by symmetric PQ capability sized according to the rated power of
means of a convex polygon (inscribed regular octagon). the connected HV/MV transformers, is attached to each PCC
Within the software implementation, both transmission and bus copy. Finally, coupling constraints are defined in the model
distribution grid expansion models can be used either indepen- to link the power exchanged by the two generators. Equality
dently, or as a combined model where both power flow formu- constraints are added for balancing active and reactive power,
lations are accommodated at the same for the transmission and if applicable. If reactive power is only supported by distribu-
distribution grids respectively. Note that when the transmission tion grid’s power flow model, reactive power exchanged with
and distribution grids are optimised together, due to the nature transmission can be bounded in a given interval.
of the power flow models, the coupling between the networks The approach for coupling the two grid levels through
will only be based on the active/reactive power injections, generators and coupling constraints has been chosen because
instead of through voltage magnitudes and/or angles as one of its versatility. As it only involves standard network com-
would expect in a typical power flow model. Considering that ponents, the complexity classes of the power flow models
typically a large number of distribution networks are connected for transmission and distribution can be changed at any time
to a single node of the meshed transmission grid (typically up without requiring adaptations to the model itself. Coupling via
to 8 medium voltage feeders per HV/MV substation), to keep a dedicated bus, on the other hand, would have required the
the tractability of the model, a novel decomposition method user to write a custom bus power balance equation for the
relying on surrogate models has been developed, which is specific pair of power flow models they intended to use.
described in the next section.
B. Decoupling heuristic
IV. T&D DECOMPOSITION USING SURROGATE MODELS Based on the idea that a better computational performance
Planning of Transmission and Distribution (T&D) networks can be expected by solving a number of small problems
can be accomplished in two ways using FlexPlan.jl. Using instead of a single, very large problem, a heuristic procedure
the combined model, introduced hereafter, the two grid levels that decouples the two network levels has been devised and
are coupled and the model is solved in a “conventional way”. implemented in order to reduce memory requirements and
The other method is the decoupling heuristic, explained in computation time of the combined transmission and distribu-
Section IV-B: it is a heuristic algorithm for decoupling the two tion network expansion planning problem. The better compu-
network levels that exploits surrogate models of distribution tational performance is obtained at the expenses of the opti-
networks to solve independent optimization problems for mality guarantee on investment decisions, since the proposed
transmission and distribution networks. The two methods are heuristic relies on the definition of an approximated surrogate
compared in Section IV-C. model for each distribution network, the purpose of which is
Transmission network data Distribution network data Distribution network
The decoupling heuristic offers two main benefits. Firstly, it in FlexPlan.jl’s repository.
2 Solver: CPLEX, with an 1·10−4 relative optimality tolerance. Hardware:
significantly reduces the amount of memory required for the desktop PC with an Intel Core i7-9700 (3 GHz, 8 cores) CPU and 64 GB
optimization process, which makes it possible to handle larger RAM.
Fig. 6. Distribution of pheuristic
P CC (t, n) − pcombined
P CC (t, n) in [−0.05, 0.05]
(for CIGRE MV networks, 8 % of the samples are outside the represented Fig. 7. Sample visualizations of planning results produced with FlexPlan.jl,
interval). When using CIGRE MV as distribution networks, the interquartile using data from the repository.
range is 0.015 (0 with IEEE33 networks) and the 1 % trimmed range is 0.41
(0.08 with IEEE33 networks).