Content uploaded by Johannes Hiry
Author content
All content in this area was uploaded by Johannes Hiry on Nov 22, 2019
Content may be subject to copyright.
25th International Conference on Electricity Distribution Madrid, 3-6 June 2019
Paper n° 1142
CIRED 2019 1/5
AUTOMATED TIME SERIES BASED GRID EXTENSIONS PLANNING USING A
COUPLED AGENT BASED SIMULATION AND GENETIC ALGORITHM APPROACH
Johannes HIRY, Chris KITTL,
Christian RÖMER, Christian REHTANZ
TU Dortmund University – Germany
johannes.hiry@tu-dortmund.de
Lars WILLMES,
Sebastian SCHIMMEYER
intulion solutions GmbH – Germany
lars.willmes@intulion.de
ABSTRACT
In recent years, the distribution grid planning process has
faced the big challenge to integrate renewable energy
sources in its planning methodology while preserving a
secure and stable provision of electricity. With the
currently observable efforts to electrify human mobility all
around the world, another new challenge arises for the
planning and operation of distribution grids. To address
these challenges and to leverage the opportunities that are
accompanied by them, new methods for the planning of
distribution grids as well as planning decision-supportive
approaches and algorithms are needed. The presented
approach contributes to the described demands by means
of a coupled approach, using both distribution grid time
series as well as a genetic algorithm to support decision
making in the planning process considering not only new
assets for grid reinforcements and extensions but also
smart-grid and operational opportunities.
INTRODUCTION
Besides the operation of their grid, one of the main tasks
of distribution grid operators (DSOs) is the suitable and
cost-efficient planning of the distribution grid under their
control. The transformation of the existing energy system
from a conventional power plant based into a renewable
energy sources based system has challenged the DSOs
planning process in recent years. [1] With the currently
globally observable efforts to electrify human mobility, the
next dare for planning and operating distribution grids is
already conceivable. Besides the conventional approach to
install new grid assets, operational flexibility provided
through battery storages, on load tap changing (OLTC)
transformers or smart market mechanisms like “traffic
light concepts” may be additionally considered in the
planning process as less cost intensive options for grid
extensions and maintenance. [2] The consideration of all
possible alternatives, including operational flexibility and
external factors, highly increases the planning effort.
Therefore, new planning methods as well as decision-
supportive approaches, considering multidimensional
dependencies as well as smart-grid mechanisms, need to
be developed. [3]
This contribution gives first insights on how an automated
decision-supportive tool for a future proof distribution grid
planning may look like. As part of the research project
“Agent.GridPlan” the functionality of the agent-based
simulation SIMONA has been extended to interact with an
existing automated genetic grid extension planning
algorithm, that uses the results from SIMONA to propose
grid extensions in case of a congestion. [4],[5]
In the first part of the paper the overall concept of the
developed coupled simulation framework and its
functionality are outlined. Afterwards, the extensions and
adjustments of the used simulation approaches are
described in detail, subsequently followed by a small
application example. In the last part, a conclusion of the
developed methodology is drawn and an outlook to further
work is given.
DEVELOPMENT OF A COUPLED
SIMULATION FRAMEWORK
To allow for the assessment of the current distribution grid
state, SIMONA has to provide specific data to a genetic
algorithm (GA). Furthermore, besides providing the data
through a specific interface and format that is
understandable by the GA, its results – proposed measures
to overcome the congestions – have to be processed and
considered for the subsequent simulation run. On the GA
part a suitable representation of the optimisation problem
needs to be found to encode the possible changes of the
power grid on one side and to allow for simple application
of genetic operators on the other side. In particular, this
involves applying said changes to the grid at an optimised
point of simulation time at which the change is actually
applied. Strictly spoken, this means that variables of the
optimisation problem consist of two components: the
actual change to the grid (“what”) and the time of
application (“when”).
Currently, the actual timing of change application is not
used since simulation time is restricted to a single year
such that timing of measures is of rather low significance.
Further development will extend simulations to multi-year
time frames where the actual placement of measures in
time will be constrained by capacity limits on financial
volume and available man power in a given time interval.
Overall coupling concept
On a high abstractive level, the overall concept of the
integrated grid evaluation and extension approach can be
structured in four main steps, depicted in Figure 1.
25th International Conference on Electricity Distribution Madrid, 3-6 June 2019
Paper n° 1142
CIRED 2019 2/5
Figure 1: Conceptual flow chart overview of the coupled
simulation framework
1. Simulation Initialisation process
During the initialisation, both, SIMONA as well as
the GA, are provided with their input data that is
held in an object relational database. It consists of
the grid asset data, scenario data, the specific agent
behavioural models as well as a set of available
options for genetic operations. Furthermore, a stop
criterion which can either be a specific amount of
simulation runs
!"#
, a convergence criterion
$
or
another, population specific measurement, is set.
SIMONA’s bottom-up structure allows for the
consideration of the whole grid in detail for all
power flow simulations without any aggregation or
simplifications.
2. Initial simulation
After the initialisation has been completed, a first
initial power flow simulation including all agent’s
behaviour is executed to check if there are any
violations in the initial grid configuration. As
several scenarios might be investigated e. g.
different penetration levels of electric vehicles
(EV) or distributed energy resources (DER) over
several years, even the initial grid configuration
might not be sufficient anymore. After the initial
power flow simulation, the calculation results are
evaluated to identify possible thermal overloads of
assets or violations of the allowed voltage range. If
none exist, the simulation will head over to the next
time frame or terminate. Otherwise, the GA grid
extension process is triggered and creates a random
initial individual to solve the grid congestion.
3. Population processing and simulation run
The third step starts with processing the optimiser
adjustments and persisting its changeset into the
results’ database. This process includes financial
valuation of the proposed individual by calculating
and persisting capital expenditures (CAPEX) and
operational expenditures (OPEX) as part of the
change set. Furthermore, resulting cost data is hand
over to the GA for being considered in the fitness
function. For performance reasons, SIMONA result
data as well as overload information are discarded
here and only the optimiser change set is persisted.
Following the changeset persisting, a steady state
power flow analysis for the predefined simulation
time window (currently one year) is performed and
corresponding time series are generated. Based on
them, the current state of the grid is valuated for
each time step by calculating potentially occurring
violations of grid constraints. The results of this
step are handed over to the GA for further
processing and application of genetic operations.
4. Fitness evaluation, genetic operations and best
solution update
The GA converts the results supplied by the
simulation to fitness values that allow for numerical
comparison of alternative solutions. Using the
fitness values tournament selection is applied to
update populations and for mating selection. New
individuals are generated by uniform crossover and
a simple mutation operator based on uniform
sampling for grid assets and on Gaussian Mutation
for timing information. The GA status is updated
accordingly and new data are exported to the
25th International Conference on Electricity Distribution Madrid, 3-6 June 2019
Paper n° 1142
CIRED 2019 3/5
runtime environment for evaluation by the
simulation.
While the steps 1 and 2 are executed only once, step 3 and
4 are repeated until the stop criterion is reached.
Upgrades and extensions of the existing time
series generation simulation
While the initial concept and design of SIMONA was to
provide the distribution grid operator with abilities to
simulate their grid, generate time series and store them into
a database for further, manual investigation, the immediate
processing of results, grid modifications on runtime and
skipping of invalid time steps hasn’t been part of the
concept yet. To allow for the aforementioned coupled
simulation several new features have to be developed.
Furthermore, as simulation runtime and performance play
a major role when simulations are carried out multiple
times, several improvements to increase the overall
performance have been addressed. The following section
provides insights into recent development and
improvements that have been made to set up the coupled
simulation. For a more detailed description of SIMONA
and its functionalities, please refer to [6].
Initialisation of the grid data, on-the-fly grid adaptions
during runtime and persisting proposed extensions
One requirement is to modify the grid model during
runtime or at least after each iteration of the genetic
algorithm. As reloading of the grid model from database
turned out to be costly in terms of simulation time, a deep
copy of the grid data is made before the initial simulation
run. This copy is hold in memory and reloaded when
adaptions proposed by the optimiser in each subsequent
simulation have to be processed.
Considering the number of iterations, the GA is executed
until its solution converges, storing and evaluating all
result data would be non-feasible. To address this
challenge, the authors reduced the amount of data to be
stored in each iteration to only the provided extension
proposals. This allows for a fast persistence, while
preserving the ability to investigate specific iterations in
detail after the coupled simulation terminates without
producing a unnecessarily big amount of unused data.
Data reduction, constraints calculation and valuation
Thanks to its modular bottom-up architecture, SIMONA is
able to provide detailed time series for each grid asset as
well as some other grid operation information. While in
the past, the overall investigation of grid asset loading data
has been the primary goal, the provided raw data is not
usable by the genetic algorithm without pre-processing,
data reduction and results valuation.
By design, the grid model in SIMONA is divided into
multiple subnets representing galvanically disjoint areas of
the grid. These subnets are represented by unique
NetAgents which carry out continuous power flow
calculations for their underlying grid. Beside this, they are
responsible for processing and persisting the
corresponding generated time series. [6] Based on the
concept of a finite state machine (FSM) the NetAgent’s
current state cannot only be observed at any time in the
simulation but also be extended by new states. To pre-
process the calculation results of each subnet, the existing
NetAgentBehaviour FSM has been extended by an
additional state that carries out the following tasks after
each time step:
1. Configuration dependent selection of relevant
calculation results
2. Computation of assets’ loading and assigning an
overload dependent penalty cost to each asset
3. Storing results in memory for further processing
4. Hand over a result vector to the genetic algorithm
In a more formal way, this new FSM state can be described
as follows.
%&"'()%*+%,+-+%&.
/"0()1+2+-+3.
45"6()4*+4,+-+45.
!"#()1+2+-+7.
(1)
Let
'
denote the set of available NetAgents,
0
the set of
pre-defined simulation timesteps,
6
the set of grid assets
under investigation and
#
the predefined number of
simulations. The calculation of the penalty costs (step 1.
and 2.) can then be written as
89:;+<+=>?@A(B>?@C1DDEFAG8+ H?@I1DDEF
D+ H?@J1DDEF
(2)
where
8
represents relative overload costs in
K F
L and
?@
the loading of the valuated asset
M
. These values are
aggregated to the >
NO1
A
P1
result vector
Q;
which is
handed over to the genetic algorithm after each run. It can
be described a
Q;(R
S
T
UVWXY+Z+[
\
Z]^
UV_`X+a+^
b
a]^ c
UV_`X+a+d
b
a]^ e
f
g
.
(3)
The simulation result vector
Q
that is handed over to the
genetic algorithm consists of the aggregated value of the
investment costs of the optimiser changeset, as well as the
aggregated penalty costs values for each simulated time
step.
Extension of the NetAgents behaviour to allow invalid
time step valuation
In SIMONA, the implemented power flow calculations are
based on a numeric Newton Raphson algorithm. While
executing this algorithm over multiple voltage levels in a
large electrical grid, non-convergent behaviour might be
observable in one or multiple time steps. In the initial
design of the simulation, the case of non-convergent power
25th International Conference on Electricity Distribution Madrid, 3-6 June 2019
Paper n° 1142
CIRED 2019 4/5
flow calculations has been considered as an invalid state.
The simulation status changed to failure instead of just
marking the corresponding time step as failed and
proceeding with the simulation. The amount of infeasible
time steps in a given period allows for a qualitative
comparison of two infeasible grid asset bases – one is
“more feasible” over another. To not only allow for a
valuation of a grid extension proposal, even with invalid
time steps, but also to send a corresponding feedback to
the genetic algorithm the existing way of executing power
flow calculations had to be extended and a second new
FSM state was introduced.
Assuming the case that a time step does not converge, the
corresponding NetAgent now informs its superior subnet
about the failure including a request to skip this time step.
As the subnet is a child of the superior net, this leads
consequently to a “fail and skip” state in the superior net
and is propagated until the highest NetAgent is informed
about the skip. At this point, the corresponding time step
is marked as failed and the corresponding cost values in
equation (2) are set to infinity. Depending on the
configuration of the GA, a predefined number of time steps
marked with infinity costs can be allowed to relax the
optimisation constraints.
Automated distribution grid extension using a
genetic algorithm
The choice to use a GA is motivated by the nonlinear and
combinatorial nature of the problem at hand. The variables
to be optimised, i.e. basically grid assets, are categorial
values drawn from a fixed set of possible values. The
showcase example is the type of an overhead transmission
line that can be used to connect two substations of the
power grid. This transmission line type defines the
electrical properties of the transmission line and its actual
cost. For the optimisation a fixed set of predefined line
types is assumed, reflecting current grid planning practice.
In GA terms the line type is used directly as a phenotypic
representation, i.e. variables are operated on directly
without further encoding. In general, standard operators
and population structures are applied to a genome built
from all line type variables.
Since the power grid contains many more potential
variables than the example line types, further assets are
added in a similar fashion. The conceptual design of the
method allows the extension to further asset classes as long
as they are supported by the grid model. This allows
optimisation of arbitrary asset classes, especially rather
exotic or special purpose ones expected to be useful in
future grid designs.
The time series nature of the simulation approach adds an
additional nonstandard feature to the optimisation
problem: any action taken on the simulated power grid,
like the exchange of a line type, is supposed to happen at a
given time within the simulation time frame. Therefore, all
optimisation variables need to have an additional time
stamp that indicates, at which simulation time the change
is applied. This time stamp is implemented as a real valued
parameter that is associated to any optimisation variable
and adapted by the optimiser. By that, the optimisation
adapts both the optimal asset type and the optimal
simulation time to apply the change. The Genetic
Algorithm uses a phenotypic representation of the time
parameter with uniform crossover and Gaussian Mutation
with fixed step size. The time stamp information is
rounded to the closest feasible value in simulation time to
match simulation requirements.
More formally, given the set of grid asset classes
6
(see
equation (1)), where each asset class has an associated set
of feasible type values
3>45A()3>45A*+hhh+3>45Ai.
for a
power grid with n components the genome of an individual
of the Genetic Optimiser is a vector of value pairs
?(
>3*+/*+hhh+3;+/;A
, where
35")3>4*A+hhh+3>4<A.
denotes
the type to be set for the i-th grid asset and
/5")1+hhh+j.
denotes the point of time in the simulation, where the
measure is taken.
The fitness function is composed of the cost associated
with the actions suggested to the simulation on the one
hand and a number of constraints that need to be met on
the other hand. Both cost and constraint values are
computed by the simulation and suitably aggregated in the
fitness function. The optimiser is capable of handling
constraints and objectives separately and does not require
constraint violations to be encoded in penalty terms added
to the objective function value. This opens the way to
optimize for strict feasibility without compromising
penalty term definitions if required.
Additional data generated by the simulation is used to
heuristically steer the search towards good solutions. Due
to rather long computation times the optimiser implements
parallel evaluation of the fitness function to speed up total
optimisation progress.
APPLICATION EXAMPLE
To investigate to correct functionality of the developed
approach a small application example has been simulated
as a first proof of concept. The grid for this case has been
selected from a large real world distribution grid [7] with
focus on traceability to ensure an easy understanding of the
GA optimisation process.
Simulation setup
The investigated grid is a real world low voltage grid
consisting of approx. 900 nodes and lines. Loads are
modelled as standard load profiles [8], DER are modelled
based on the descriptions in [6]. Simulations are carried
out for one day in January with an hourly resolution. Six
different options for line extensions and eight different
options to extend the transformer capacity were available
to the GA. The chosen scenario leads to a high share of
additional PV penetration in the grid.
Simulation Results
As a first result, Figure 2 demonstrates the fundamental
operating principle of the developed approach. The
25th International Conference on Electricity Distribution Madrid, 3-6 June 2019
Paper n° 1142
CIRED 2019 5/5
abscissa shows the wall clock time from the computation
on a single processor machine, the right ordinate and red
graph denotes aggregated congestion and asset
investments costs in Euro, while the left ordinate and blue
graph denotes the valuated grid congestion. The
congestion being a constraint is required to be less than or
equal to 0 for feasible solutions, and costs are intended to
be minimised.
Figure 2: Optimiser convergence in the application example
As apparent in Figure 2, a convergence of grid congestion
and grid extension costs in the course of an optimisation
run takes place. On a first glimpse the convergence implies
a failure of the optimisation method, because it neither
achieves feasible nor optimal values, since the constraint
value is strictly greater than 0 and the cost is actually
increasing where it should be minimised. On closer
inspection it turns out though, that after an initial drop of
the congestion value there remains one transformer in the
simulated grid that is consistently overloaded by a small
amount. A further review of the available asset options and
the simulation setup reviled that, with the provided asset
changeset, the overloading on one transformer can’t be
fully removed, but is in an acceptable range, because a
slight overloading only occurs in a few hours of the
simulated time series. Still the optimiser is capable of
reducing the actual overload during optimisation, although
it applies actions that, from an engineering standpoint,
would be a bad idea. In fact, the optimiser replaces a
couple of other transformers and lines by alternatives with
different impedances. Consequentially, the congestion
indicator value is decreased slightly at the expense of
rather big increases in total cost. Although the technical
result achieved by the optimiser might be of non-optimal
engineering value, it still shows that both the method and
its prototypical implementation work as supposed.
CONCLUSION AND OUTLOOK
In this contribution, a methodology to couple an agent-
based grid simulation with a genetic algorithm for grid
extension planning has been presented. Its fundamental
functionality for has been demonstrated in a very small
application example. As a next step, the developed
methodology will be refined and extended in a way to
supply additional measures for the optimiser to improve
the results from an engineering point of view.
Furthermore, as the demonstrated application example
represents a comparably small application example from a
large real world distribution grid, the developed approach
is going to be applied to the whole grid over multiple
voltage levels.
Acknowledgments
The authors gratefully thank Westnetz GmbH for
supporting the presented research by granting access to the
mentioned real-life grid model during their participation in
the research project Agent.GridPlan. It has been funded by
the European Fund for Regional Development under grant
agreement number EU-1-1-006.
REFERENCES
[1] Deutsche Energie-Agentur GmbH (editor), 2012,
“Ausbau- und Innovationsbedarf der Stromverteilnetze
in Deutschland bis 2030.”, Berlin, available:
https://www.dena.de/fileadmin/dena/Dokumente/Pdf/
9100_dena-Verteilnetzstudie_Abschlussbericht.pdf
[2] J. Kays, A. Seack, U. Häger, 2016, “Consideration of
innovative distribution grid operation concepts in the
planning process”, Proceedings of 2016 IEEE PES
Innovative Smart Grid Technologies Conference
Europe (ISGT-Europe), Ljubljana, DOI:
10.1109/ISGTEurope.2016.7856331.
[3] L. Jendernalik, D. Giavarra, C. Engels, A. Maier, 2016,
„Resilient and comprehensive networks planning is
key to a successful integration of renewable energy
resources”, CIRED Workshop 2016, Helsinki
[4] Thomas Bäck, D.B Fogel, Z Michalewicz, 1997
“Handbook of Evolutionary Computation”, CRC Press
[5] L. Jendernalik, C. Engels, J. Hiry, C. Kittl and C.
Rehtanz, 2017, A Holistic Planning Approach:
Enhacement of the grid expansion using the flexibility
of network participants, CIRED - Open Access
Proceedings Journal, vol. 2017, no. 1, pp. 2312-2315.
[6] J. Kays, C. Rehtanz, 2016, “Planning process for
distribution grids based on flexibly generated time
series considering RES, DSM and storages”, IET
Generation, Transmission & Distribution (IET GTD),
vol. 10, 3405-3412. DOI: 10.1049/iet-gtd.2015.0825.
[7] C. Kittl, J. Hiry, C. Wagner, C. Rehtanz, C. Engels,
2019, "Large scale agent based simulation of
distribution grid loading and its practical application",
Proceedings of the 25th International Conference on
Electricity Distribution (CIRED), Madrid
[8] H. Meier, C. Fünfgeld, T. Adam, B. Schlieferdecker,
1999, “Repräsentative VDEW-Lastprofile”, available:
https://www.bdew.de/media/documents/1999_Reprae
sentative-VDEW-Lastprofile.pdf
