0% found this document useful (0 votes)
12 views

SmartenergynetworkdigitaltwinsFindingsfromaUK-baseddemonstrator project

The document discusses a Digital Twin framework for electrical distribution systems, implemented in a UK-based Smart Energy Network Demonstrator (SEND) microgrid. It highlights the challenges of integrating imperfect model and measurement data, demonstrating that measurement uncertainties significantly affect Digital Twin outputs, and estimates potential benefits such as reduced curtailment. The findings emphasize the need for robust data assimilation processes to enhance decision-making and operational efficiency in distribution networks.

Uploaded by

Miftah
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views

SmartenergynetworkdigitaltwinsFindingsfromaUK-baseddemonstrator project

The document discusses a Digital Twin framework for electrical distribution systems, implemented in a UK-based Smart Energy Network Demonstrator (SEND) microgrid. It highlights the challenges of integrating imperfect model and measurement data, demonstrating that measurement uncertainties significantly affect Digital Twin outputs, and estimates potential benefits such as reduced curtailment. The findings emphasize the need for robust data assimilation processes to enhance decision-making and operational efficiency in distribution networks.

Uploaded by

Miftah
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 12

Electrical Power and Energy Systems 162 (2024) 110302

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems


journal homepage: www.elsevier.com/locate/ijepes

Smart energy network digital twins: Findings from a UK-based demonstrator


project
Matthew Deakin a ,∗, Marta Vanin b,c , Zhong Fan d , Dirk Van Hertem b,c
a Newcastle University, Newcastle-upon-Tyne, UK
b
Electrical engineering department (ESAT) of KU Leuven, Heverlee, Belgium
c
EnergyVille, Genk, Belgium
d
University of Exeter, Exeter, and was previously academic director of SEND at Keele University, UK

ARTICLE INFO ABSTRACT

Keywords: Digital Twins promise to deliver a step-change in distribution system operations and planning, but there are
Digital twins few real-world examples that explore the challenges of combining imperfect model and measurement data,
Digitalization and then use these as the basis for subsequent analysis. In this work we propose a Digital Twin framework
Distribution system state estimation
for electrical distribution systems and implement that framework on the Smart Energy Network Demonstrator
Microgrid
microgrid in the UK. The data and software implementation are made available open-source, and consist of
a network model, power meter measurements, and unbalanced power flow-based algorithms. Measurement
and network uncertainties are shown to have a substantial impact on the quality of Digital Twin outputs.
The potential benefits of a dynamic export limit and voltage control are estimated using the Digital Twin,
using simulated measurements to address data quality challenges, with results showing curtailment for an
exemplar day could be reduced by 56%. Power meter data and a network model are shown to be necessary for
developing algorithms that enable decision-making that is robust to real-world uncertainties, with possibilities
and challenges of Digital Twin development clearly demonstrated.

1. Introduction management tasks, such as for substation transformers [4], or condi-


tion monitoring of medium voltage (MV) overhead line systems [5].
The digitalization of energy systems continues to gather pace, with Industry is also showing a strong appetite for these approaches, with
new system architectures promising to increase utilization, improve utilities developing increasingly ambitious plans for the development
resilience, and support the societal transition towards net zero. These of Digital Twins of their service areas [6].
architectures can incorporate real-time monitoring and dispatch in Despite this strong outlook, publications reporting real-life expe-
electricity distribution networks, enabling functionalities such as dy- riences of developing Digital Twins are uncommon, partly due to
namic operating envelopes, flexibility markets, or peer-to-peer energy researchers’ limited access to necessary utility data [7]. Typically,
markets. Digitalization therefore has a central role in transforming a Digital Twin has two steps: a data assimilation stage, reconciling
passive distribution networks into an active and responsive distribution measurement data and system model (e.g., through formulation as
systems that can host much higher levels of low carbon technologies. a Distribution System State Estimation (DSSE) problem) [8,9]; and
This proliferation of system data is leading both industry and an analysis phase, which combines the processed data with other
academia to imagine the potential role of Digital Twins in future dis- computational methods to support operational or planning decision-
tribution systems. Whilst definitions of Digital Twins vary substantially making tasks. Despite DSSE having been studied for several decades,
between sectors [1], for distribution systems they typically combine academic case studies are typically based on synthetic test cases which
a network model with measurements to support decision making by fail to show few of the complexities of real networks [10] and lack the
network operators and planners. For example, in [2], the authors necessary power and voltage measurement data [11,12]. For example,
a collection of 128 real LV feeders from the UK were developed in [13],
propose an online method for assigning Volt/VAr controller set points
but a lack of appropriate smart meter data meant that simulated
for low voltage (LV) distribution systems. Similarly, a real time Digital-
active power profiles had to be provided (if DSSE is required, artifi-
Twin based method is demonstrated for a multi-energy network in [3].
cial test cases can be created by recording voltages at measurement
Digital twins have also been proposed for power distribution asset

∗ Corresponding author.
E-mail address: [email protected] (M. Deakin).

https://doi.org/10.1016/j.ijepes.2024.110302
Received 17 November 2023; Received in revised form 17 June 2024; Accepted 9 October 2024
Available online 30 October 2024
0142-0615/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

buses and adding artificial noise [14]). In other cases, power flows
and voltages from real networks are made public, but usage of this
data can be hampered by a lack of corresponding network model and
unspecified measurement accuracy [15]. In contrast, there are many
methods for supporting decision making through control- or market-
based means [16,17], but decision-making algorithms typically do not
consider the data assimilation process, and so would not nominally
be considered a Digital Twin. Such analysis methods may thus fail
to fully exploit available information in the decision making process,
or make unrealistic assumptions as to data accuracy and robustness. Fig. 1. The proposed Digital Twin framework aims to combine network models and
telemetry with DSSE and simulation algorithms. Section 3 discusses the physical
There is therefore a need for test cases which demonstrate practical
system, network model, online measurements, and data historian; Section 4 discusses
imperfections in distribution network and measurement data: we show the network model and DSSE; Section 5 discusses distribution system simulation and
that these are valuable not only for developing ‘whole’ Digital Twin decision making activities.
applications, but also for DSSE on its own.
There is a small but growing literature on experiences with real data
for use with Digital Twins or DSSE. Several papers report findings and 2. A digital twin framework for distribution network applications
experiences using real data that cannot be shared (e.g., for privacy or
commercial reasons). For example, [18] present the results of DSSE on Definitions of Digital Twins vary significantly between sectors and
a Swiss distribution network, highlighting how mismatches between applications [1]. For the purposes of this work, we consider a dis-
real and digital line parameters cause their state estimator to flag tribution system Digital Twin to be a software representation of the
‘normal’ current measurements as gross data errors. Großet al. [19] physical system that is characterized by the five elements illustrated in
evaluate the sensitivity of DSSE outputs to the pseudomeasurements Fig. 1. Functionally, the Digital Twin acts as an interface for support-
required in an MV network. Others explore the impact of measure- ing decision making processes undertaken by control room operators
ment configuration and error on DSSE results [20], or how incorrect and system planners, with the aim of supporting both planning and
transformer impedances affect the development of Digital Twins [11]. operational decisions.
Finally, a small number of test cases feature both network data and real The Digital Twin takes input data from the physical system in the
power data from the network’s smart meters, such as the interconnected form of a network model and online measurements. The network model
LV systems in [21] and the unbalanced MV/LV test system of [22]. includes all information required to run a power flow, i.e., network
However, to the authors’ knowledge, no prior distribution test cases topology, impedances (cables types and lengths), transformer models,
feature a network model together with both power and voltage data. and load/generator connectivity. Measurement data from power meters
This is because energy (power) data are easier to obtain, as utilities and other data acquisition systems can be used online and stored in a
collect them for billing purposes, while voltage measurements are central data historian for subsequent interrogation. The network model,
typically not stored. However, errors in the twin’s network model can
online measurements and data historian can be used for simulation
only be identified if voltage measurements are available, and syntheti-
of the distribution system in unrealized circumstances (using a dis-
cally created voltage measurements do not necessarily capture real-life
tribution system simulator), or for estimating the state of the system
complexities. Given the role of extensive demand-side electrification in
given noisy measurements (using DSSE algorithms). The main differ-
reaching net zero, the need for representative test systems which more
ence between the Digital Twin and a simulation set-up rests in data
closely represent the network and monitoring data available to DNOs
assimilation: through the access to field measurements, the Digital Twin
in contemporary systems is both significant and timely.
can utilize the most up-to-date load and network data for planning and
The contribution of this paper is to explore the opportunities and
operational tasks, enabling significantly increased accuracy. Table 1
challenges of developing Digital Twins in power distribution systems
lists the software used to carry out these functionalities.
by proposing a Digital Twin framework for a real-world MV/LV dis-
tribution system, then acquiring the models and data alongside a
2.1. Comparing digital twins with existing practice
software implementation to realize this twin. Opportunities to provide
inputs for decision-making for system owners are demonstrated as a
In transmission networks, data acquisition and assimilation are part
clear potential benefit of the Digital Twin, whilst also highlighting the
of state estimation workflows since the 1970s, and control rooms
risks if an insufficient model and data fidelity cannot be reconciled
receive updated information on the system’s (steady) state every few
to create a consistent system representation. The network model [23]
minutes. Transmission networks are in general fully observable, and
and measurement data [24] are made available open-source, with a
measurement semantics are well understood. Therefore, Digital Twins
workbook for reproducing results of this work at
could be seen as an extension of current control room practice, rather
than the development of a new operating approach.
https://github.com/deakinmt/DSSE_SEND.
In contrast, distribution network practices vary hugely depending
The paper is structured as follows. Section 2 proposes a Digital Twin on geography, regulation, and policies. It is thus challenging to pro-
for power distribution networks, highlighting the crucial requirements vide a clear overview of the state of the art and compare it with
for DSSE within the Digital Twin for data assimilation. Section 3 the proposed Digital Twin. However, in many distribution systems,
provides a detailed explanation of the Smart Energy Network Demon- developing a Digital Twin would represent a substantial change in the
strator (SEND) microgrid, on which the Digital Twin framework is utility operating practice. For example, the dominant ‘fit-and-forget’
implemented. Section 4 discusses the challenges encountered working model of distribution operations and planning only uses data on an
with the real data, showing how the present meter configuration and ad-hoc basis, and there is no control room as there are no remote
network model require improvement to reach the standards needed for actions the DNO will be expecting to undertake. The SEND system is
effective data assimilation. The benefits of a dynamic voltage control somewhat unusual in this regard (as a demonstrator project) as it has
scheme are estimated in Section 5 to highlight a potential use-case of a relatively high level of monitoring. Nevertheless, even within this
the Digital Twin, albeit using the Digital Twin in a synthetic setup high-value demonstrator, data assimilation is not provided, diminishing
(i.e., using a combination of real and simulated ‘measurements’). In the ability to identify errors in models and data, or to be able to use
Section 6 we survey potential future directions and research gaps for an optimal estimate of the network state from DSSE. In summary,
Digital Twins, before offering conclusions in Section 7. prior practise from DNOs contrast strongly with the proposed Digital

2
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

Table 1
Tools used to realize the SEND digital twin.
Digital twin module Tool or format
Online measurements Siemens SICAM P855
Data historian Siemens Distributed Energy Optimization (DEOP); cached in .csv format
Network model OpenDSS .dss format
Distribution system simulator OpenDSS (nonlinear power flow); Julia (linear power flow)
Distribution System State Estimation (DSSE) PowerModelsDistribution StateEstimation.jl [14]

Fig. 2. The SEND virtual private wire system (or ‘microgrid’) interconnects with the
wider distribution network at four points.

Twin concept for distribution systems. The Digital Twin allows planners
and control engineers to exploit a common historic best-view of the
network, based on models and data, with the aim of providing a
step-change in accuracy and granularity of information available for
decision-making.
Typical challenges that distribution network Digital Twins need to
address include integrating disparate databases (e.g., combining GIS
information directly with parameters such as line construction codes);
providing seamless interfaces to third-party information (e.g., AMI
data) as and when required; and improved information recording for
critical parameters such as switch states or off-load transformer tap
positions. Additional organizational bottlenecks and more informed
engineering practices are discussed in [10], such as enhanced mea-
surement semantics, better integration of data sources, and systematic
methods for validating electrical network models.
Fig. 3. Augmented single line diagram of the MV network of the SEND virtual private
wire network showing the network, distributed energy resources and location and
3. The smart energy network demonstrator voltage level of power meters. For conciseness, LV substation transformers are not
shown.

The Smart Energy Network Demonstrator (SEND) [25] is located


at the Keele University campus in the West Midlands of the UK, and Table 2
Summary of SEND microgrid parameters.
aims to provide an energy system ‘living laboratory’ [25]. From the
Parameter Value
distribution system perspective, this includes substantial investment in
Number of MV buses 33
power quality metering equipment and the associated data collection
Total no. nodes 174
platform. As shown in Fig. 2, the SEND electrical system operates Export limit (buses 13-1, 13-2) 1.578 MVA
as a virtual private wire system, i.e., it is fully owned and operated Export limit (buses 1, 2) 0 MVA
by the university, rather than by the local network operator. As a Total AC generation capacity 7.1 MVA
result, the university is responsible for all network and measurement Minimum demand 0.24 MVA
Baseload demand 4.03 MVA
data. The system is therefore an ideal sandbox to explore challenges
Total MV cable length 13.1 km
and opportunities for Digital Twins. This section presents the network MV upper voltage limit 1.06 pu
model and measurement data for SEND that were used to implement LV upper voltage limit 1.10 pu
the Digital Twin framework (as described in Section 2). MV lower voltage limit 0.94 pu
LV lower voltage limit 0.94 pu

3.1. Network topology and model

The augmented single line diagram in Fig. 3 shows the electrical microgrid can export. However, even from this point of connection,
network topology, the location of distributed energy resources, and the voltage constraints mean that the export at the point of common
voltage and location of the power meters. A summary of key network coupling must remain less than 1.578 MW. As the total connected
parameters is reported in Table 2, showing that the network peak generation capacity is greater than this value, the site is subject to an
generation is much higher than the baseload demand. The topology of export limit, with protection in-place so that the generators will trip
the MV network is weakly meshed, with LV substations (not shown) if there is a greater export than this value for a period of more than
feeding demands connected radially. Within the LV networks, only the 10 s [26]. To avoid tripping of the protection, an alarm is therefore
LV substation busbar is modeled. raised after one second by the energy management system, then after a
There are four potential points of coupling to the upstream network further four seconds generation is curtailed by the energy management
(as also shown in Fig. 2): bus 1, 2, 13-1, or 13-2. The network is nor- system. It is therefore not uncommon to see oscillatory, ‘hunting’
mally operated with only bus 13-2 connected to the main distribution behavior in PV output during high solar output, as the generation is
network, as this is the point of common coupling (PCC) for which the reduced due to over-export, slowly ramped up over the course of a

3
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

Table 3 Table 5
Virtual power plant assets, together representing approximately 400 kVA demand-side Summary of SEND electrical telemetry parameters.
response. Parameter Value
Parameter Number
Meter type Siemens SICAM P855 [27]
Air conditioning unit 1 No. MV meters 7
Air handling units 24 No. LV meters 15
Pump 5 double, 11 single Synchronization accuracy ±25 ms
Fan coil units 20 Update frequency Twice per minute
Electric boiler 3 Measurands Line–line voltages magnitudes [V], phase currents
Smart plug 95 [A], apparent power [kVA], total real power [kW],
Immersion heater 10 total reactive power [kVAr], three-phase power
Chiller 7 factor.
Car chargers 11 dual chargers
Air handling unit circuit pumps 1 double
Variable temperature circuit pumps 1 double

Table 4
Dates in 2022 of data scraped from the SEND DEOP data historian API.
Date Weekday Date Weekday
13th May Friday 16th Sept. Friday
14th May Saturday 17th Sept. Saturday
15th July Friday 11th Nov. Friday
16th July Saturday 12th Nov. Saturday

few minutes, and then reduced once again due to the over-export (as
considered in more detail in Section 5).

3.1.1. Distributed energy resources Fig. 4. Voltages, currents and power measurements for the solar PV generator
Fig. 3 shows that in addition to the generators, there are a number connected close to bus 30 for 15th July 2022.
of other distributed energy resources (DERs). Together, these DERs can
be scheduled to reduce both the total cost of energy supplied to the
SEND campus and overall campus carbon intensity. For example, when
surplus generation is predicted, the site’s gas boilers can be turned off
and the campus water heating demand is instead met via three 0.5 MW
electric boilers.
In addition to the boilers, approximately 400 kVA of demand side
response technologies are split across a wide range of different demand
classes, summarized in Table 3. It is interesting to note that, even
though there are more than 150 assets, together they provide less
flexibility than just a single large electric boiler. Nevertheless, they still
represent a significant opportunity to shift demand temporally to make
use of renewable energy that would otherwise be spilled.

3.2. Telemetry

As shown in Fig. 3, power meters are spread throughout the net-


work. Table 5 summarizes their settings. At present, there are twenty Fig. 5. Range, interquartile range and median MV line–line voltage measurements
across all buses for 15th July 2022.
two monitoring devices connected to the distribution network (as SEND
continues to develop, this number is growing). These meters can mea-
sure a wide range of relevant power quality phenomena [27], and are
configured to record rms line voltages, phase currents, three-phase real Section 4.1). On this date, there is also solar curtailment, as can be
and reactive power, each at a 30 s interval (Table 2). The recorded observed in the power injection shown in Fig. 4(c) (on this day, if the
data is transmitted to a data collection platform (Siemens’ Distributed export limit is crossed, the solar PV output is reduced to 0.9 MW).
Energy Optimization Platform, DEOP), where data can be visualized or Naturally, voltages also change across the network through the day.
queried from an API for further analysis. Eight days of data have been This can be visualized by considering how the range, interquartile
scraped from the API and provided in csv format. These dates repre- range and median voltage varies, as plotted in Figs. 5 and 6 for the
sent the second Friday and Saturday of the month every other month MV and LV networks. By comparison with Fig. 4(a), it can be observed
from May to November 2022, and are given in Table 4. These dates that over the course of the day that voltage magnitudes follow the
cover weekday and weekends, and represent between them periods of pattern of solar voltages, but that the spread is much greater for the
low, moderate and high system demand (undergraduate students are LV voltages. This increase in spread is due to different turns ratios for
vacant from the campus in June and July). the transformers in the network (discussed in Section 4).
Fig. 4 shows the measured voltages, currents and powers for the
solar bus on 15th July 2022. The line–line voltages are high, and exceed 3.2.1. Modeling measurement uncertainty
the nominal MV voltage limit of 1.06 pu for short periods of time. The measurement uncertainty and valid measurement range of the
Currents and powers, as expected, increase through the morning and power meters vary for different measurands, as shown in Table 6.
reduce through the evening, although the measured phase-𝑎 current is Assuming the meter’s rated voltage 𝑈 Rt d. is the equivalent to the
consistently zero (this and other data quality issues are addressed in per-unit operating voltage of the network, these meters will capture

4
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

effective Digital Twin. Firstly, Section 4.1 discusses the quality of raw
data from power meters, showing how a number of data quality issues
require preprocessing. Challenges in network model development are
outlined in Section 4.2, considering data inconsistency and incomplete
component information. In particular, inaccurate MV/LV transformer
models are shown to have a significant detrimental impact on the
Digital Twin’s performance. Finally, challenges around power meter
configurations are outlined in Section 4.3. We show that, even assuming
a perfect network model and noiseless measurements, meter paucity
and inadequate measurement semantics make it impossible to recover
the state of the system reliably. A bespoke, heuristic DSSE approach is
proposed to address these issues, but a clear need for improved data
quality is seen (with regards to measurements and network model) to
enable real-world DSSE-based decision making.

Fig. 6. Range, interquartile range and median LV line–line voltage measurements


4.1. Measurement data quality
across all buses for 15th July 2022.
The power meters installed in SEND (Section 3.2) are designed to
Table 6 collect 30 s rms measurements continuously to a prespecified measure-
Measurement relative errors for the SICAM P855 (from the datasheet [27]). ment accuracy (Table 6). However, by inspection of the recorded data,
Measurand Operational uncertainty Valid range it is clear that the measured data does not meet that accuracy in all
Line–line voltage 0.5% 0% to 120% of 𝑈 Rt d. cases.
Phase currents 0.2% 20% to 200% of 𝐼 Rt d. Some data quality issues can be observed in the raw plotted data
Real power 1% 20% to 200% of 𝐼 Rt d. in Fig. 7. In some cases, measurement errors are trivial, such as meters
Reactive power 2% 20% to 200% of 𝐼 Rt d.
that are ‘stuck’ on a single value for each measurement, like that in
Fig. 7(a) (an issue also seen for 𝐼a in Fig. 4(b)). Other meters show
mixed problems, such as in Fig. 7(b): one of the measurement values is
voltages during normal operations with operational uncertainty 𝜖 𝑈 of fixed, whereas the two other measurements follow a stepped pattern.
0.5% (voltage bounds on MV networks in the UK are between 0.94 and This occurs if the device only records a change if some threshold is
1.06 pu [28], and so during normal operation network voltages are less met (for example when the measurand has changed by some fixed
than 120% of 𝑈 Rt d. ). percentage). Furthermore, individual time stamps may present a trivial
Current and power measurements have relative error of 0.2% for gross error or missing value, as shown in Fig. 7(c). Finally, a small
currents between 20% and 200% of the meter’s rated current 𝐼 Rt d. . number of meters have been installed but are yet to be configured,
These flows are often much less than 20% of 𝐼 Rt d. , and so the accuracy and so return non-numerical values (NaNs). It is worth noting that new
is not specified at those ratings. Here, it is assumed that the absolute meters returning large errors is often due to incorrect firmware settings
current measurement error 𝜖𝑖,𝐼 𝜙 does not deteriorate as the current (firmware can typically be updated rather than requiring a new meter
reduces below this value (subscripts indicate indexing, with 𝑖 the bus to be installed).
index and 𝜙 the phase index). That is, for the current measurement on Finally, some phenomena that might be associated with data quality
bus 𝑖 and phase 𝜙, |𝐼| ̂ 𝑖, 𝜙 , the absolute current operational uncertainty issues were identified as measurement tolerance issues. An example is
does not drop below the value at 20% of the rated current for the device that of the solar plant meter measuring small power demands overnight
at bus 𝑖, 𝐼𝑖Rt d. , i.e., (not plotted). Such event could suggest that a load is connected but
{ } ended unreported in the network data. However, as overnight currents
𝜖𝑖,𝐼 𝜙 = 0.2% × max |𝐼| ̂ 𝑖, 𝜙 , 0.2 × 𝐼 Rt d. . (1)
𝑖 are <20% of the rated one, the solar meter presents a large relative error
We note that in academic DSSE papers, the valid region of the opera- (see (1)), which leads to powers with an incorrect sense (i.e., ‘wrong
tional uncertainty is typically neglected, leading to very optimistic as- sign’ for the active power). It is interesting to note that sense for the
sumptions as to current and power measurement uncertainty at loading power that is consistently incorrect indicates bias in the measurements,
below 20% as compared to (1). contrasting with the common assumption of zero-mean Gaussian mea-
surement error (which would imply both positive and negative errors
4. Assessing digital twin adequacy via distribution system state during times when there is no generator export).
estimation
4.2. Network model development
For a Digital Twin to fulfill its role as a decision-making tool,
measurement data and network models need to be assimilated to create The Digital Twin’s network model was built considering the type
a meaningful, non-trivial and self-consistent system representation. In and length of MV cables connecting the substations, with sequence
the context of the SEND Digital Twin, the voltage outputs of DSSE components being used to create cable impedances. In general, the data
and distribution system simulations (power flow) can be compared to for creating this model (obtained in a previous project) was not fully
historic measurement data to perform this validation. A functioning consistent, and some elements (system topology, cable types) required
Digital Twin will have DSSE outputs that are within tolerance of meter confirmation either through review of GIS information or via discussion
operational uncertainties, with large differences between measured with university estates. While we assume their impact to be modest,
and calculated variables being associated with low confidence in the potential inaccuracies in those elements’ values remain.
Digital Twin’s performance. In the case of inadequate performance, A more substantial challenge rests in the estimates of the tap posi-
Digital Twins can still be studied as synthetic, ‘virtual’ Digital Twins tions of the MV/LV transformer models. The transformers turns ratios
(considered in Section 5). can influence the output voltage by double-digit percentages across
In this section, we show that the data quality and network modeling their full range, and the incorrect choice of tap positions in the Digital
accuracy in the present configuration need to be improved for the real- Twin may lead to substantial simulation and estimation errors. Trans-
world system at SEND to achieve the high standard required to be an former models are based on provided datasheets that report winding

5
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

Fig. 7. Three types of errors identified by inspection.

types, impedances, and nominal voltages for primary and secondary


windings. However, it is known that manual tap positions changes may
go unreported [10], and so actual tap positions may be very different
from their nominal positions.
To consider the accuracy of taps as nominally provided, a non-
linear unbalanced power flow was run, with the SEND network model’s
PCC (ss13-1) serving as the slack bus. To convert from line to phase
voltages, slack bus voltage angles are assumed 120◦ apart and then
magnitudes determined from the measured line voltages. Loads are
assigned according to their measured power or according to a nominal
split of unaccounted power at the PCC. The power flow solution with
nominal tap positions is shown in Fig. 8(a), with each of the three
points corresponding to the meter’s three line voltages. Large discrep-
ancies can be observed between measured and simulated voltages: for Fig. 8. Comparing simulated versus measured line–line voltages, for 12.14 pm on
example, simulated voltages at ‘ss14’1 are much higher than measured September 17th, before and after updating MV/LV tap positions. Each color represents
values, whilst those of ‘ss18’ are lower. While not plotted, the pattern three measurements at a single measurement point. Updated taps lead to a much closer
fit (the dashed line represents the line 𝑦 = 𝑥, a perfect fit would lie exactly on this
of Fig. 8(a) is consistent over time, and is largely independent of
line).
assumptions on network loading.
It is possible to reduce this error via a trial-and-error heuristic by
manually increasing or decreasing the MV/LV tap positions to yield
are particularly useful for DSSE as they implicitly include information
a reasonable fit, with the improvement in fit shown in Fig. 8(b).
on the phase angle between local current and voltage phasors, and
This resulted in the simulated and measured line voltages becoming
the current angle has a significant impact on voltage drops across
much more aligned. In principle, a field inspection could confirm the
a multi-phase impedance. Furthermore, as the DSSE state typically
transformer type and position at all locations; however, this is resource-
includes per-phase power variables, providing only their sum (one
intensive. We note that self-correcting network models are a desirable
measurement per three-phase load) reduces observability as compared
feature for future-proof Digital Twins [10]. However, due to meter
to a measurement strategy utilizing three per-phase measurements
paucity, the corrections could not be automated, i.e., a human-in-the-
loop was required for this exercise. ‘Augmented’ DSSE examples exist, per load. Therefore, contemporary unbalanced DSSE algorithms are
with tap ratios included as problem variables alongside conventional typically reliant on per-phase power measurements [14].
states (e.g., [29]). These augmented DSSE require full observability, a Lastly, the power meters are only placed in a limited subset of
property which does not hold in SEND (as described in Section 4.3). nodes, resulting in an unobservable system. In contrast to power me-
ter configuration issues, this scarcity of measurement devices is a
4.3. DSSE and observability well-known challenge for DSSE [9].
In principle, all three of the above issues could be addressed. Power
In general, reliable DSSE requires both a very good network model meters can be configured to provide per-phase voltages and powers,
and measurement devices which are appropriately placed and config- and in some situations utilities are requesting this as a requirement
ured. With regards to the latter, there are three specific challenges for the meters that are being installed (e.g., [30]). Power meters can
in the SEND measurement system today. Firstly, power meters are also be retrofitted to existing substations to improve observability
configured to report line voltages 𝑈 LL rather than phase voltages; line (note that in SEND, no measurement devices are available other than
voltages do not have a common reference or neutral, and so zero at the substations highlighted in Fig. 2). However, in this instance,
sequence voltages cannot be detected. As phase voltages are typically it proved to be too costly to increase the number of power meters
considered to be the output state of DSSE, unbalanced DSSE algorithms or reconfigure the meters and data acquisition system (e.g., to en-
usually assume the availability of phase voltage measurements. DSSE able per-phase power measurements or phase voltages). Furthermore,
implementations must therefore be adjusted to take line voltages as whilst there may be a very wide set of potential meter configurations
input (but nonetheless without a reference voltage, will be unable to (e.g., choice of measurands, accuracy ratings, percentage substation
uniquely specify voltages with respect to ground). coverage) which can support DSSE from a theoretical perspective, it
Secondly, the meters are configured to report only total power may only be viable for commercial DSSE solutions to use the most
flows, rather than per-phase powers. Per-phase power measurements reliable and well-understood subset of those meter configurations. Ad-
justing meter settings and software implementations post-deployment
may be uneconomical, and so utilities and practitioners should strive to
1
Substations are referred to according to their reference in the network specify measurement requirements adequately prior to meter, firmware
model [23], as compared to the simpler single line diagram of Fig. 3. and DSSE deployment.

6
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

4.3.1. DSSE approach and results be returned that is valid for a set observed nodes (without requiring
To run DSSE, we use the open-source PowerModelsDistributionSta- any assumptions of the statistics of the demand or voltages at the
teEstimation.jl package [14]. This tool is attractive for the SEND system unobserved nodes).
as it can solve under-determined state estimation problems without The full DSSE formulation is given in the Appendix, and is a
relying on pseudomeasurements, and because the optimization-based non-convex quadratically constrained quadratic programming (QCQP)
framework is easily customizable, enabling bespoke measurement con- problem. Results for one timestep are visualized in Fig. 9, both with the
straints such those necessary to include line voltage measurements 𝑈 LL . nominal and updated MV/LV tap positions (as described in Section 4.2).
This ease-of-prototyping is possible thanks to the underlying automatic These figures plot the residual, calculated as the difference between the
differentiation toolbox JuMP [31]. measured line voltage and the estimated voltage as calculated via DSSE.
As is conventional for DSSE problems, measurement errors are Whilst the true value of state variables cannot be known (due to in-
assumed to be independent and normally distributed, with standard evitable measurement noise), approximate bounds can be inferred from
deviation one third of the maximum error, meter tolerances (Table 6). The results from DSSE for the model with
( )
0.005 × 𝑈̂ 𝑖,LLpq nominal MV/LV tap positions (Fig. 9(a)) indicate that the measurement
𝑈𝑖,LLpq ∼  𝑈̂ 𝑖,LLpq ; , (2) and network models are inconsistent. This cannot be explained due
3
to potential errors that may be introduced by the heuristic (3)–using
where 𝑈̂ 𝑖,LLpq is the line voltage measurement at bus 𝑖 measured from the voltage–power sensitivity (7), introduced later in Section 5.1 and
phase p to phase q. parameterized in Table 7, a 0.15 pu voltage deviation would require
Without per-phase current or power flow measurement, the system an injection of more than 20 MW, much greater than the maximum
would be unobservable even if all nodes had phase voltage and three- demand of the network (see Table 2). A number of buses are com-
phase power measurements. Therefore, for the purposes of this work, pletely unobservable, as can be identified due to their residual having
we combine current magnitudes |𝐼| ̂ 𝑖,𝜙 and total powers 𝑃̂ t ot into a
𝑖 a value that is numerically zero (seven buses have all three voltage
heuristic composite per-phase power measurement 𝑃̂𝑖, 𝜙 for phase 𝜙 for residuals less than 6 × 10−5 ). These buses have numerical value zero,
the 𝑖th bus, as in those under-observed subsystems the state estimator will have
|𝐼|
̂ 𝑖,𝜙 sufficient degrees of freedom to assign measurement–variable pairs
𝑃̂𝑖, 𝜙 = 𝑃̂𝑖t ot ∑ . (3)
as having identical values whilst still maintaining consistency of the
𝜙∈{𝑎,𝑏,𝑐} |𝐼|𝑖,𝜙
̂
power flow equations.
An equivalent identity holds for a per-phase reactive power. Note that, The output of DSSE following the tap position update is plotted in
as discussed previously in Section 4.3, ideally in future work the meters Fig. 9(b), with the residuals having decreased significantly as compared
will be updated to enable the per-phase powers to be used directly, to the nominal MV/LV tap positions. It can be observed that changing
rather than requiring the use of the heuristic (3); as such, simulations
the tap positions did not result in a change in residuals for unobservable
conducted in Section 5 assume per-phase measurements. (Active and
nodes (as expected). The residuals are still too variable to consider a
reactive powers measurements are used rather than using the measured
good DSSE output. Nevertheless, they confirm in a statistical sense that
power factor and apparent power because of the non-linear relation
the updated network model improves the Digital Twin’s data quality.
between power factor and reactive power leading to poor measurement
error propagation). The uncertainty of (3) is not known, but is required
5. Case study: Using the digital twin for estimating benefits of a
for DSSE. This uncertainty is based on (1) and is estimated according
dynamic export limit
to
{ }
𝑃
𝜖𝑖,𝜙 = 0.01 × 𝑈𝑖Rt d. × max |𝐼|
̂ 𝑖,𝜙 , 0.2 × 𝐼 Rt d. ,
𝑖 (4) The SEND network is subject to a static maximum export limit, set
by the utility as a result of voltage congestion, following the UK stan-
leading per-phase powers to be distributed as
( 𝑃 )
dard G100 [26]. To meet this standard, the SEND energy management
𝜖𝑖,𝜙
system automatically reduces solar power output by several MW if the
𝑃𝑖,𝜙 ∼  𝑃̂𝑖,𝜙 ; . (5)
3 power threshold is breached for a set time period. When this power
reduction occurs, the energy management system acts to increase de-
An equivalent distribution exists for per-phase reactive power mea-
mand from electric boilers and other flexible demand sources, then the
surements (with the exception that the operational tolerance of 1% is
solar export ramps back towards the available power export. However,
replaced with 2%, as in Table 6).
in general, the maximum solar generation is large in comparison to the
There are two general approaches to perform DSSE in unobservable
systems: providing pseudomeasurements at unobserved power injec- energy that can be absorbed by the flexible loads and electric boilers.
tions [9], resulting in a ‘pseudo-observable’ system, or solving the In this section, we use the SEND Digital Twin to consider a potential
underdetermined state estimation problem directly. Pseudomeasure- solution to address these high curtailment volumes. In particular, a dy-
ments have been the dominant approach historically, as they enable namic export limit is considered, with voltage control also provided by
the use of classical, well-established state estimation approaches that setting a non-unity power factor for the solar generator. In Section 5.1
are common in transmission system applications. However, several we explain the physical mechanism by which this approach would
authors have recently considered alternative approaches that question reduce curtailment, with the role of the Digital Twin in calculating
pseudomeasurements’ value, as there is a risk they can introduce signif- the potential benefits of this approach described in Section 5.2. The
icant noise and compromise estimation quality [32,33]. Proposed DSSE analysis is based on a synthetic test case, with results presented for
methods that avoid pseudomeasurements include the use of alternative September 17th (an exemplar day with significant curtailment) in
statistical metrics [33], matrix completion [32], and the identification Section 5.3.
of ‘‘observable islands’’ through the use of pseudo-inverses of singular
matrices [34]. In this work, we choose the latter approach for DSSE 5.1. Reducing curtailment with a dynamic export limit and reactive power
for two reasons. Firstly, whilst [32,33] present several advantages control
over pseudomeasurements, including reduced historical data require-
ments, they still rely on statistical assumptions on non-monitored users Changing the SEND connection agreement to have a dynamic export
for results to be valid; there is, however limited information at the limit would mean that exports will only be limited if relevant state
non-observed nodes within the SEND system on which to base those as- parameters have reached their operational limits at a given time,
sumptions. Secondly, as we demonstrate in Section 4.3.1, there are non- rather than having only been determined for a conservative, worst-
trivial observable islands—i.e., a non-trivial DSSE result can therefore case scenario (as is the case for the existing static export limit). As the

7
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

As per common practice (e.g., [26]), it is assumed that opera-


tional uncertainty is accounted for in the dynamic export limit control
schemes, given the safety-critical function of the controller. Thus, when
voltages approach their limits, there will be some curtailment 𝑃 SF due
to the tolerance with respect to voltage estimates 𝜖 (in %) as
𝑃 SF = 𝜖 𝛥𝑃 Solar, 1% , (9)

where 𝛥𝑃 Solar, 1% is the power injection–voltage rate (7). This means


that techniques that reduce 𝜖 (e.g., through sensor fusion in DSSE) can
reduce the curtailment that would be calculated using raw measure-
ments directly. If only a single voltage measurement is available then
the tolerance 𝜖 is identical to its operational uncertainty 𝜖 𝑈 (as reported
in Table 6).

5.2. Using the digital twin for network and curtailment modeling

There could be a large outlay for SEND to request a dynamic export


limit, due to complex network studies for the utility and any addi-
tional systems that may be required (e.g., development of utility-side
DSSE [36]). Given that there is a growing database of historic conges-
tion and curtailment events, the Digital Twin is an ideal candidate to
calculate the estimated benefit. With respect to Fig. 1, SEND system
planners can use data from the data historian to consider curtailment
volumes, use DSSE to provide the best estimate of the historic system
state, then use distribution system simulation to consider a counterfac-
tual with a dynamic export limit and voltage control enabled. In turn,
the reduced curtailment volumes can be used to calculate additional
revenue for the university and reduction in carbon emissions.
As shown in Section 4, the real world network model and measure-
ment data are not fully consistent, and so a synthetic test case is used
for this case study. A heuristic process combines the network model
Fig. 9. Line voltage residuals following DSSE. with power and voltage measurements to create an unbalanced system
with the slack voltage and load and generation powers all specified (in-
cluding allocating load at nodes without measurements). Specifically,
SEND system is the only customer which exports power on its feeder for nodes with measured real and reactive powers, these powers were
in the network (Fig. 2), it is assumed that there will not be any further equally split amongst phases. The system residual power (the difference
voltage rise downstream of it. Therefore, this voltage-based dynamic between power injected and total power across measured substations)
export limit is equivalent to ensuring that voltage magnitudes do not is then allocated among loads and phases accordingly, with those indi-
exceed their limits at the SEND PCC or within its network. As can be vidual allocated powers uniformly distributed between 0% and 200%
seen in Figs. 5 and 6, measured voltages indicate significant headroom of the mean per-load average. To reduce computational burden, every
during many hours of the day, albeit with slight overvoltages during fourth measurement is used (i.e., at a two minute temporal resolution).
late morning and afternoon hours. Once those loads have been allocated, synthetic voltage magnitudes
Given the fact that voltage congestion is more likely during periods and power measurements are then determined by running a power
of high solar output, it is proposed that the power factor of the solar in-
flow and adding white noise to those synthetic measurements according
verter is also adjusted to reduce the impact of generating active power.
to the tolerances given in Table 6 (with standard deviation one third
The benefit of changing generator power factor can be illustrated by
of the total error). To avoid the challenges described in Section 4.3,
considering how a reactive power injection 𝑄Solar and active power
the synthetic measurements that are captured consist of phase volt-
injection 𝑃 Solar result in a voltage rise 𝛥𝑈 due to positive sequence
impedance 𝑍 = 𝑅 + 𝚥𝑋 as (approximately) [35] ages and per-phase powers at all load and generation buses. Under
these conditions, the network is observable and the DSSE can be used
𝛥𝑈 ≈ 𝑃 Solar 𝑅 + 𝑄Solar 𝑋 . (6) reliably.
Reducing the power factor to draw 𝑄Solar
(i.e., an increasingly negative The amount of solar curtailment at a given time instance is not
𝑄Solar ) results in an increased active power 𝑃 Solar that can be injected by estimated within the data historian DEOP. Physics-based curtailment
the solar plant for the same voltage rise. For example, the solar injection estimation is a complex task (requiring, for example, panel temper-
that results in a 1% voltage rise 𝛥𝑃 Solar, 1% can be calculated from (6) ature, direct and diffuse irradiance, shading, and solar panel electri-
considering the generator power factor PFSolar as cal parameters). For the purposes of this work, a heuristic approach
1% is therefore used to estimate curtailment. The GB-wide solar profile
𝛥𝑃 Solar, 1% = , (7)
𝑅 − 𝑋 𝑐 PF from [37] is normalized, scaled and offset according to the installed

1 − (PFSolar )2 capacity at the solar generator, to estimate the potential generation
𝑐 PF = . (8)
PFSolar during periods of known curtailment. The curtailment is then estimated
Although (6) is only an approximation, in practice (7) provides good as the difference between the potential and measured power output
accuracy (as considered in Section 5.2), with estimates of 𝑅 and 𝑋 (with any spurious negative values assigned a value of zero). The
determined via power flow linearization techniques. resulting profile was validated by inspection for the date considered.

8
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

Fig. 10. Maximum network voltage, including measurement error bars.

5.2.1. Implementation approach


A linear power flow formulation is used to calculate changes to Fig. 11. The solar generation model (‘Solar model’), modeled curtailment (‘Est. Curt.’),
the system voltage magnitudes in SEND 𝑈 SEND as compared to the and additional potential solar generation 𝑃 Solar, Max for three dynamic curtailment
schemes.
estimated voltage magnitudes from DSSE 𝑈 DSSE due to real and reactive
injections at the solar generation 𝑃 Solar , 𝑄Solar ,
[ Solar ] Table 7
𝑃 Power injection–voltage sensitivity 𝛥𝑃 Solar, 1% (from (7)) and corresponding safety factor
𝑈 SEND = 𝑀 + 𝑈 DSSE , (10)
𝑄Solar 𝑃 SF (from (9)) for tolerance 𝜖 = 0.5%.
Power factor Sensitivity 𝛥𝑃 Solar, 1% , MW Safety factor 𝑃 SF , MW
where 𝑀 ∈ R3𝑛×2 is the linear sensitivity matrix for a system of 𝑛 three
1.0 1.640 0.820
phase buses. This approach was chosen as it is conceptually simple, 0.9 2.355 1.178
avoiding a need for iterative power flow calculations. The linear matrix
𝑀 was created using the approach outlined in [38] and was found
to have very good accuracy, with the relative error in the 2-norm of
voltage changes found to be between 1% and 9.7% or less for per-phase • A second scheme uses a power factor of 0.9 (labeled ‘Q Control’).
real or reactive injections of 200 kVA at any node in the network. This scheme is used to illustrate the additional generation possible
due to the addition of an example voltage control scheme, also
The maximum possible solar injection 𝑃 Solar, Max for a given power
assuming that operational uncertainty 𝜖 in DSSE outputs is small.
factor will happen at the smallest solar injection 𝑃 Solar for which any
• Finally, a third dynamic scheme considers unity power factor
bus reaches the upper voltage limit 𝑈 + . This can be calculated from
generation, but considers the impact of operational uncertainty
(10) as 𝜖 on the additional generation (labeled ‘Conservative’). This uses
( [ ])
1 as a voltage for calculations the smaller of the raw measurements
𝑃 Solar, Max = min(𝑈 + − 𝑈 DSSE )∕𝑒 𝑀 PF (11)
−𝑐 and DSSE output plus the operational tolerance 𝜖 𝑈 of 0.5%.
where ∕𝑒 denotes elementwise division and coefficient 𝑐 PF maps real to The ‘Conservative’ model results in the smallest additional generation
reactive powers for a given fixed power factor, as in (8). 𝑃 Solar, Max , and the ‘Q Control’ approach the largest. In addition, on this
figure the estimated total solar and estimated curtailment are plotted.
From Fig. 11, it can be seen that curtailment takes up a substantial
5.3. Results fraction of the potential solar generation on this day. Nevertheless, the
utilization of the network could clearly be increased substantially—in
The maximum (synthetic) MV measured voltage in the SEND system the hours before 8 am, more than 2 MW of additional power could be
for 17th September is shown in Fig. 10 alongside the tolerance 𝜖 𝑈 of generated from the site for all dynamic export schemes, and after 2 pm,
± 0.5%. It can be seen that for short periods, there are over-voltages. there is sufficient capacity even for the conservative scheme to mitigate
A dynamic export limit would avoid these issues (although, it is noted all present curtailment. There are some negative values of 𝑃 Solar, Max in
that power quality standards typically allow voltages to stray outside the late morning period as the measured maximum voltage is greater
of nominal bounds by a small amount, so long as this is only for a short than 1.06 pu for a small number of time periods.
Note that the differences between the three curtailment reduc-
duration).
tion schemes can be explained succinctly by considering the voltage–
We use (11) to calculate the additional solar generation 𝑃 Solar, Max
curtailment rate 𝛥𝑃 Solar, 1% and curtailment safety factor 𝑃 SF . These are
that could be obtained with three different control schemes, instead of calculated with respect to voltage rise at solar bus in Table 7. The
static curtailment. Results are shown in Fig. 11. power injection–voltage sensitivity 𝛥𝑃 Solar, 1% increases the power that
can be injected by 44%, resulting in ‘Q Control’ being (close to) a factor
• A first scheme considers a generator exporting with unity power of 1.44 times that of ‘DSSE + Lin. Mod.’, as can be observed by-eye.
factor (labeled ‘DSSE + Lin. Mod.’). This model illustrates the The curtailment safety factor 𝑃 SF has value of 0.82 MW, and so the
additional benefit of moving to a dynamic limit without addi- ‘Conservative’ scheme has a constant factor close to this value less than
tional voltage control. It is assumed that the DSSE process means can be exported as compared to ‘DSSE + Lin. Mod.’.
that the operational uncertainty 𝜖 can be assumed to be small Fig. 12 plots the estimated energy curtailed for the three models. It
(e.g., that 𝜖 ≪ 𝜖 𝑈 ). can be seen from this figure that all methods can reduce the curtailment

9
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

6.0.0.3. Maturation of partially-observable DSSE methodologies. In tradi-


tional power system applications, the main benefit of state estimation is
to provide a systematic approach for recovering the system state using
a self-consistent physical system model, and to do so with the highest
possible accuracy. In distribution systems, obtaining full observability
for all buses and line flows is typically not cost-effective. In the physical
DSSE calculations described in Section 4.3, our approach has been
to consider only the estimation of the state via an underdetermined
formulation, and so the state is returned only as a non-null value in
an observable subset of nodes. However, in a real distribution system,
it may be the case that the operator does need an estimate of the full
system state. To address this, either the network operator can install
further measurement devices, or use pseudomeasurements. However,
without good historic data on which to base the pseudomeasurements,
the uncertainty of estimated state could be large.2 Therefore, even
though the state estimator would technically be observable from a
Fig. 12. Bar chart showing the total estimated curtailment against three control mathematical perspective, the true uncertainty can be so large as
approaches that utilize a dynamic export limit, as compared to the existing static export to be operationally meaningless. Therefore, where pseudomeasure-
limit.
ments are necessary, efforts should be made to ensure these are as
accurate as possible (e.g., by aggregating lower fidelity smart meter
measurements).
on this day by a significant amount. The approach based on unity Furthermore, DSSE methods tailored to consider partially observ-
power factor control results in a reduction in curtailment of 4.2 MWh, able systems could be implemented. In particular, developing solutions
although if the utility insists on conservative state estimation, this to automatically identify observable sub-areas appears a promising
reduces the output by 2.69 MWh. In contrast, the ‘Q Control’ model re- research direction. Within those sub-areas reliable state and parameter
duces curtailment volumes by 5.82 MWh. For this latter case, an energy estimation is possible.
price of $100/MWh and grid carbon intensity of 400 kgCO2e/MWh
6.0.0.4. Development of test systems for digital twins and DSSE. We note
implies potential to increase revenues by $582 and reduce emissions
that other publicly available distribution system test cases (both real-
by 2.33 TCO2e for this day.
world and synthetic) provide power flows at loads and generators, but
not voltage measurements, and so are too simplistic for Digital Twin
6. Discussion: Towards Digital Twins applications. This is because state estimation and parameter identifi-
cation tasks require voltages to ensure the system is over-determined,
and to capture the complexity that is associated with unknown errors
The development of the network models and collection of data in network models and measurement uncertainty. Given this lack of
for the proposed system has resulted in a ‘synthetic’ Digital Twin of test cases, authors must make heuristic assumptions as to the quality
the SEND distribution system, with more work required to develop of metered data, meter placement, and typically assume a known and
a self-consistent, fully functioning twin. Much of the software and perfect network model. Additionally, in those cases, noise is added to
hardware implementation now exists for the system, but there are the power flow results, yielding unreproducible residuals (unless the noise
outstanding issues which highlight general challenges that distribution is provided together with the results). Furthermore, the structure of
system Digital Twins will face. meter errors may be very complex–measurement errors can drift with
time, leading to correlated errors. DSSE problems never have a known
6.0.0.1. Defining the scope of the digital twin framework. The Digital ground-truth, and so real-world validation methods have a different
Twin framework proposed in Section 2 may need to be adapted to flavor to the results output from synthetic test cases. Finally, a future
account for additional scope depending on its application. For example, test system could also consider LV systems more comprehensively,
the curtailment model was considered a component of the ‘data histo- using GIS information and consumer smart meter data, as DSSE in LV
rian’ block; however, in applications with many customers (e.g., in the networks is typically heavily reliant on those measurements.
setting of domestic solar PV dynamic operating limits), the forecasting
of customer-side DERs becomes a substantial task in itself, with output 7. Conclusions
depending on solar PV capacities, orientation, and shading. Closer
connections with other energy sectors, such as with electric vehicles Digital Twins for distribution systems are set to become a
or heating systems, would also require substantial modeling efforts to widespread architecture as energy system digitalization strategies reach
achieve adequate performance. fruition in coming years. They promise to enable new operational and
planning approaches, with potential to maximize the value of data
6.0.0.2. Automated network model identification and correction. As from newly commissioned measurement devices in a self-consistent
shown in Fig. 1, the network model is influenced by DSSE and network representation of the physical energy network. In this work, we have
simulation, as the network model must be validated for real-world proposed a Digital Twin framework, then implemented this for the
applications. It has been proposed that DSSE can also be augmented to SEND demonstration campus at Keele University, UK. As a publicly-
include the identification of network parameters directly, particularly funded demonstrator project, it was possible to provide open network
with respect to tap settings [29,39]. The approach of these works models, measurement data and algorithms for exploring non-idealities
of real-world Digital Twin components. This contrasts with distribution
is to consider how residuals over a number of time periods can be
system test systems today, which do not consider where power meters
minimized, and so tap settings are reported jointly with the system
state. Given the requirement for power meters to be reconfigured to
allow reasonable state estimation, this was not considered, but it is 2
We note that estimating good pseudomeasurements for individual loads,
noted that these approaches have been shown to be very effective in e.g., households, is particularly challenging. This task may be easier for
real-world applications [40]. aggregated loads.

10
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

might realistically be installed, their tolerance, or in non-trivial net- original implementation (15) is added for the SEND case. The DSSE
work modeling errors, each of which can have substantial effects on problem is then solved using Ipopt [41].
subsequent analysis. DSSE is performed in the AC-rectangular (ACR) variable space, with
State estimation has been proposed as an integral part of the Digital phase voltage phasors 𝐔 ∈ C𝑛×3 and lifted power variables 𝐒 ∈ C𝑛×3×3
Twin architecture. The SEND measurement data and network model represented within the optimization in real and imaginary components,
underlines the necessity of both correctly configured meters and a
network model with a very good accuracy to enable an adequate state 𝐔𝑖 = 𝐔Re Im
𝑖 + 𝑗𝐔𝑖 ∀𝑖 ∈  , (12)
estimation output. We hope that other researchers can use the model 𝐒𝑖𝑗 = 𝐏𝑖𝑗 + 𝑗𝐐𝑖𝑗 ∀(𝑖, 𝑗) ∈  ∪  , R
(13)
and data for development and validation of new algorithms to address
the challenges seen in this project. For example, future work could where  is the set of buses, , R
are the sets of branches in forward
include integration of DSSE with model identification functionality to and reverse orientation, and bold math font indicating a variable which
address challenges seen in this work around identifying tap settings of can be indexed only via the first element (e.g., 𝐒𝑖𝑗 will return a 3 × 3
MV/LV transformers. matrix). Loads and generators are defined via
The Digital Twin has been used to explore the potential for dy-
𝐒𝑔𝑘 , 𝐒𝑑𝑘 ∀𝑘 ∈  ∪  , (14)
namic export limits and voltage control to minimize solar curtailment.
However, there is huge potential for a range of further applications where ,  are the sets of generators and loads respectively.
both in microgrid and grid-connected systems, such as the scheduling
SEND voltage measurements are line voltage magnitudes |𝑈 LL |,
of virtual power plant assets, developing consumer dynamic operating
which are not nominally included in the ACR variable space. These
envelopes, or preventive and corrective actions during stress events. We
can therefore be incorporated into the formulation via the equality
conclude that distribution system Digital Twins could be a cornerstone
constraint
technology for network operators, and that utilities, academia and
2
industry must work together to develop these as a holistic and flexible |𝑈𝑖,LLpq | = |𝑈𝑖, p − 𝑈𝑖, q |2
(15)
tool for the wide range of distribution grid characteristics seen around ∀𝑖 ↦ 𝑚 ∈ |𝑈 |pq , ∀pq ∈ {𝑎𝑏, 𝑏𝑐 , 𝑐 𝑎}.
the world.
As discussed in Section 4, the raw data from the power meters return
CRediT authorship contribution statement only total three-phase powers, rather than per-phase quantities. These
are split according to the heuristic (3), and can be linked directly to
Matthew Deakin: Writing – review & editing, Writing – original demand or generator variables (14).
draft, Methodology, Funding acquisition, Formal analysis, Data cura- The DSSE objective function is the weighted sum of the least squares
tion, Conceptualization. Marta Vanin: Writing – review & editing, of residuals,
Writing – original draft, Software, Methodology, Investigation, Formal ∑
minimize 𝜌𝑚 , (16)
analysis, Conceptualization. Zhong Fan: Writing – review & editing,
𝑚∈
Resources, Funding acquisition. Dirk Van Hertem: Writing – review &
editing, Resources, Funding acquisition. where weighted residuals 𝜌𝑚 are defined as
( )2
𝑥𝑚 − 𝑧𝑚
Declaration of competing interest 𝜌𝑚 = ∀𝑚 ∈  . (17)
𝜎𝑚2

The authors have no interests to declare. Here, 𝑥𝑚 is the system variable corresponding to the measured value 𝑧𝑚 ,
with 𝜎𝑚 representing the confidence on the measurement accuracy. The
Acknowledgments raw measurements are line voltages and (heuristic) per-phase powers
(used in Section 4.3), where the synthetic measurements used in the
The authors are grateful for extensive support in accessing data case study are phase voltage and per-phase powers for the synthetic
and developing the Digital Twin from Ashley Dean, Matthew Dean case study (used in Section 5.3).
and Ian Shaw at SEND, Keele University. The authors also thank Dr The power flow equations (Kirchhoff’s current law and generalized
Frederik Geth from GridQube for helpful discussions on observability Ohm’s law) complete the DSSE model:
and state estimation. The Smart Energy Network Demonstrator project
(ref: 32R16P00706) is part-funded through the European Regional De- 𝐒𝑖𝑗 = 𝐔𝑖 𝐔𝐻 𝑐 𝐻 𝐻 𝐻
𝑖 (𝐘𝑖𝑗 + 𝐘𝑖𝑗 ) − 𝐔𝑖 𝐔𝑗 𝐘𝑖𝑗 ∀(𝑖, 𝑗) ∈ , (18)
velopment Fund (ERDF) as part of the England 2014 to 2020 European 𝐒𝑗 𝑖 = 𝐔𝑗 𝐔𝐻 + 𝐘𝑐𝑗𝑖 )𝐻 − 𝐔𝐻 𝐻
∀(𝑖, 𝑗) ∈ , (19)
𝑗 (𝐘𝑖𝑗 𝑖 𝐔𝑗 𝐘𝑖𝑗
Structural and Investment Funds (ESIF) Growth Programme. It is also ∑ 𝑔 ∑ ∑
receiving funds from the Department for Business, Energy and Indus- 𝐒𝑘 − 𝐒𝑑𝑘 = diag(𝐒𝑖𝑗 ) ∀𝑖 ∈  . (20)
trial Strategy (BEIS), UK. M. Deakin was supported by the Centre for 𝑘∈𝑖 𝑘∈𝑖 (𝑖,𝑗)∈𝑖 ∪𝑖R

Postdoctoral Development in Infrastructure, Cities and Energy (C-DICE)


In the above, (⋅)𝐻 indicates the conjugate transpose, 𝐘𝑖𝑗 represents the
programme (funded by the Research England Development Fund) and
series admittance of the branch’s 𝛱-model (see [42,43]), whereas 𝐘𝑐𝑖𝑗
the Royal Academy of Engineering, UK under the Research Fellowship
is the shunt admittance. 𝑖 , 𝑖 are the sets of generators and loads
programme.
connected to bus 𝑖. The form of (18)–(20) has first been proposed
Appendix. DSSE formulation as a multiphase branch flow model in [44]. Note that, if the source
voltage and branch power flows are known, then all voltage phasors
In contrast to conventional DSSE, which is typically based on non- can be determined by solving the non-linear equations (18)–(20). This
linear least squares minimization solved with Gauss–Newton methods, is in direct analogy to (balanced) transmission system applications,
the DSSE algorithms used in this work are based on (equivalent) where phasors can be determined without phasor measurement units
optimal power flow-based formulations, with the general approach (PMUs) at buses, up to an arbitrary reference phase angle. Again,
described in [14]. As described in Section 4.3.1, the main advan- as in those transmission system applications, a lack of synchrophasor
tage of this approach is that it is easily customizable and can solve measurements necessitates the solution of a non-linear state estimation
under-determined DSSE problems conveniently. For completeness, we problem. Future demonstrators incorporating distributed PMUs could
reproduce the full formulation here, noting that with respect to the incorporate those measurements using the same formulation (18)–(20).

11
M. Deakin et al. International Journal of Electrical Power and Energy Systems 162 (2024) 110302

Data availability [22] Sandell S, Bjerkehagen D, Birkeland B, Sperstad IB. Dataset for a Norwegian
medium and low voltage power distribution system with industrial loads. Data
Brief 2023;48:109121.
The data and code are available, as outlined in the introduction of
[23] Deakin M, Fan Z, Dean A, Dean M, Shaw I. Network model for a smart energy
the paper. network digital twin. 2023, http://dx.doi.org/10.25405/data.ncl.21618342.v1.
[24] Deakin M, Vanin M, Fan Z, Dean A, Dean M, Shaw I. Telemetry data for a
smart energy network digital twin. 2023, http://dx.doi.org/10.25405/data.ncl.
References 21618273.v1.
[25] Fan Z, et al. The role of ‘living laboratories’ in accelerating the energy system
[1] Dalibor M, et al. A cross-domain systematic mapping study on software decarbonization. Energy Rep 2022;8:11 858–64.
engineering for digital twins. J Syst Softw 2022;193:111361. [26] Energy Networks Association (ENA). Engineering recommendation g100, issue 2
[2] Gui Y, et al. Automatic voltage regulation application for PV inverters in low- 2022 amendment 2: Technical requirements for customers’ export and import
voltage distribution grids–a digital twin approach. Int J Electr Power Energy Syst limitation schemes. 2023, https://www.ena-eng.org/ena-docs/. (Accessed 13
2023;149:109022. August 2023).
[3] Hang T. Digital twins of multiple energy networks based on real-time simulation [27] Siemens. Class S power quality instrument and power monitoring device SICAM
using holomorphic embedding method, part II: Data-driven simulation. Int J P850/P855 7KG85xx v02.23/v02.60 device manual. 2020.
Electr Power Energy Syst 2023;153:109325. [28] UK Government. The electicity safety, quality and continuity regulations 2002.
[4] Moutis P, Alizadeh-Mousavi O. Digital twin of distribution power transformer for 2002, https://www.legislation.gov.uk/uksi/2002/2665. (Accessed 13 December
real-time monitoring of medium voltage from low voltage measurements. IEEE 2022).
Trans Power Deliv 2020;36(4):1952–63. [29] Nanchian S, Majumdar A, Pal BC. Ordinal optimization technique for three-phase
[5] Gauce D, et al. Application of digital twin in medium-voltage overhead distribution network state estimation including discrete variables. IEEE Trans
distribution network inspection. Remote Sens 2023;15(2):489. Sustain Energy 2017;8(4):1528–35.
[6] Northern Powergrid. Digitalization strategy and action plan. 2022, https://www. [30] Northern Powergrid. Nps/007/021–technical specification for secondary distribu-
northernpowergrid.com/downloads/12243, (Accessed 18 August 2023). tion substation monitoring systems. 2022, https://www.northernpowergrid.com/
[7] Mateo Domingo C, et al. A reference network model for large-scale distribu- document-library. (Accessed 25 August 23).
tion planning with automatic street map generation. IEEE Trans Power Syst [31] Dunning I, Huchette J, Lubin M. JuMP: A modeling language for mathematical
2011;26(1):190–7. optimization. SIAM Rev 2017;59(2):295–320.
[8] Abur A, Gómez-Expósito A. Power system state estimation: theory and [32] Donti PL, Liu Y, Schmitt AJ, Bernstein A, Yang R, Zhang Y. Matrix completion for
implementation. CRC Press, CRC Press; 2004. low-observability voltage estimation. IEEE Trans Smart Grid 2020;11(3):2520–30.
[9] Primadianto A, Lu C. A review on distribution system state estimation. IEEE [33] Pau M, Pegoraro PA. WLS-based state estimation for unobservable distribu-
Trans Power Syst 2017;32(5):3875–83. tion grids through allocation factors evaluation. IEEE Trans Instrum Meas
[10] Geth F, Vanin M, Van Hertem D. Data quality challenges in existing distribution 2024;73:1–13.
network datasets. In: Proc. CIRED conf. Rome, Italy. 2023, p. 1–6. [34] Krause O, Martin D, Lehnhoff S. Under-determined WLMS state estimation. In:
[11] Lave M, Reno MJ, Peppanen J. Distribution system parameter and topology IEEE PES Asia-Pacific power and energy engineering conf. 2015, p. 1–6.
estimation applied to resolve low-voltage circuits on three real distribution [35] Turitsyn K, Sulc P, Backhaus S, Chertkov M. Options for control of reactive power
feeders. IEEE Trans Sustain Energy 2019;10(3):1585–92. by distributed photovoltaic generators. Proc IEEE 2011;99(6):1063–73.
[12] Vanin M, Geth F, D’hulst R, Van Hertem D. Combined unbalanced distribution [36] Energy Networks Association. Guidelines for actively managing voltage levels
system state and line impedance matrix estimation. Int J Electr Power Energy associated with the connection of a single distributed generation plant. 2012.
Syst 2023;151:109155. [37] Wilson G, Godfrey N. Electrical half hourly raw and cleaned datasets for Great
[13] Rigoni V, et al. Representative residential LV feeders: A case study for the North Britain from 2009-11-05. 2021, http://dx.doi.org/10.5281/zenodo.4573715.
West of England. IEEE Trans Power Syst 2016;31(1):348–60. [38] Bernstein A, et al. Load flow in multiphase distribution networks: Exis-
[14] Vanin M, Van Acker T, D’hulst R, Van Hertem D. A framework for constrained tence, uniqueness, non-singularity and linear models. IEEE Trans Power Syst
static state estimation in unbalanced distribution networks. IEEE Trans Power 2018;33(6):5832–43.
Syst 2022;37(3):2075–85. [39] Korres G, Katsikas P, Contaxis G. Transformer tap setting observability in state
[15] Deakin M. Annual half-hourly real and reactive power flows for 171 primary estimation. IEEE Trans Power Syst 2004;19(2):699–706.
electrical distribution circuits in the north of England. 2023, http://dx.doi.org/ [40] Vanin M, Hertem DV. The role of state estimation in the improvement of
10.25405/data.ncl.22047758. low voltage distribution network models. 2022, https://tinyurl.com/yc34xnjy.
[16] Alam M. Allocation of dynamic operating envelopes in distribution networks: (Accessed 22 August 23).
Technical and equitable perspectives. IEEE Trans Sustain Energy 2023. [41] Wächter A, Biegler L. On the implementation of an interior-point filter
[17] Zhang Z, Li R, Li F. A novel peer-to-peer local electricity market for joint trading line-search algorithm for large-scale nonlinear programming. Math Program
of energy and uncertainty. IEEE Trans Smart Grid 2019;11(2):1205–15. 2006;106(1):25–57.
[18] Zanni L, et al. PMU-based linear state estimation of lausanne subtransmission [42] Fobes DM, Claeys S, Geth F, Coffrin C. Powermodelsdistribution.jl: An open-
network: Experimental validation. Electr Power Syst Res 2020;189:106649. source framework for exploring distribution power flow formulations. Electr
[19] Groß D, et al. Evaluation of a three-phase distribution system state estimation Power Syst Res 2020;189:106664.
for operational use in a real medium voltage grid. In: IEEE PES ISGT Europe. [43] Claeys S, Deconinck G, Geth F. Decomposition of n-winding transformers for
2019, p. 1–5. unbalanced optimal power flow. IET Gener Transm Distrib 2020;14(24):5961–9.
[20] Früh H, et al. Evaluation of a three-phase distribution system state estimation [44] Gan L, Low SH. Convex relaxations and linear approximation for optimal
for operational use in a real low voltage grid. In: Proc. of the 9th renewable power flow in multiphase radial networks. In: 2014 power systems computation
power generation conf. Online; 2021, p. 125–30. conference. IEEE; 2014, p. 1–9.
[21] Koirala A, Suárez-Ramón L, Mohamed B, Arboleya P. Non-synthetic European
low voltage test system. Int J Electr Power Energy Syst 2020;118:105712.

12

You might also like