1300 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO.

5, MAY 2015

A Scalable Data-Driven Monitoring Approach

for Distribution Systems
Mohsen Ferdowsi, Student Member, IEEE, Andrea Benigni, Member, IEEE, Artur Löwen, Behzad Zargar,
Antonello Monti, Senior Member, IEEE, and Ferdinanda Ponci, Senior Member, IEEE

Abstract— This paper introduces a new data-driven bottom-up DSOs are able to take corrective actions and plan for
monitoring approach for distribution systems. In this approach, component upgrades and grid-extension investments.
local estimations of the subsections into which the system is Although monitoring is already widely used in High
split are performed independently, thus leading to a scalable
architecture. The monitoring approach is focused only on the Voltage (HV) transmission grids, applying the same
estimation of voltage magnitude rather than the complete state monitoring approach and equipment for distribution networks
of the system. This reduces the measurement requirements sig- is not feasible. The main reason is that the overall Medium
nificantly, thus addressing economical and technical concerns for Voltage (MV) and Low Voltage (LV) electricity grids are
existing systems, while staying open to accommodating further much larger than HV grids in terms of the number of nodes
incremental improvements in the available data and data quality.
The estimation of each section is realized via an artificial neural and the total length of the lines. For example, the Italian
network (ANN), for which a set of parameterizations is available HV grid operated by the Italian transmission system operator
to cope with different operating conditions. The estimation Terna is approximately 63 500-km long, whereas the total
convergence is achieved even with relatively few measurements, length of MV and LV grids managed by Enel, Italy’s largest
although accuracy varies depending on the available measure- DSO, is around 1 125 000 km [2], [3]. As a result, applying the
ments. At the Medium Voltage (MV) level, where reconfiguration
is common, a configuration identification unit chooses the right approach used in HV as it is to MV/LV grids would require so
ANN, the one trained for the actual network configuration. The much measurement and monitoring equipment that it would be
estimation process is computationally simple and can be executed economically prohibitive. To tackle this problem, monitoring
on low-cost hardware, as demonstrated in this paper by the systems should use few measurements, where the computation
implementation on a BeagleBone Black board. To demonstrate should be light and incremental deployment must be possible.
the concept, a prototype and a laboratory setup have been
developed. The experimental test results are presented both for an Meeting these requirements may result in lesser performance,
Low Voltage distribution system and an MV distribution system. which, however, at the moment may actually satisfy the
needs of the DSO. Performance can later be improved by
Index Terms— Artificial neural networks (ANNs), hierarchical
systems, power distribution, power system monitoring, state incrementally upgrading the original measurement system.
estimation. In recent years, the approaches to state estimation for
distribution networks presented in the literature mainly rely
I. I NTRODUCTION on variations of the weighted least squares (WLS) method.
Singh et al. [4] present an optimization method for meter

D ISTRIBUTION system operators (DSOs) are responsible

for the secure operation and maintenance of their
systems, as well as for planning and performing necessary
placement that aims at satisfying specified target accuracy for
the state estimates. Pau et al. [5] describe a novel approach to
the integration of synchrophasor measurements. Liu et al. [6]
upgrades and extensions to ensure their long-term ability to present the use of a heterogeneous set of measurements,
meet the demand [1]. To fulfill these duties properly and including smart meters, PMUs, substation measurements, and
efficiently, DSOs need to be aware of the ongoing situation pseudo and virtual measurements, together with an optimiza-
within their distribution grids via monitoring systems. tion method to identify the best measurement configuration.
In particular, using the data collected by the monitoring In [7], it is shown that correlation of measurements and
systems, DSOs can identify issues, such as outages, excessive pseudomeasurements may have significant impact on the
voltage deviations, and overloads. With this information, estimation accuracy in some cases and must, therefore,
be accounted for. Concerning dynamic measurements in
Manuscript received June 20, 2014; revised January 21, 2015; accepted
January 22, 2015. Date of current version April 3, 2015. The Associate Editor distribution grids, Castello et al. [8] define a new algorithm
coordinating the review process was Dr. Carlo Muscas. for synchrophasor measurement that satisfies both P and M
M. Ferdowsi, A. Löwen, B. Zargar, A. Monti, and F. Ponci are accuracy classes as defined by the IEEE Standard C37.118.1
with the Institute for Automation of Complex Power Systems, E.ON
Energy Research Center, RWTH Aachen University, Aachen 52062, and the results apply particularly well to distribution grids.
Germany (e-mail: [email protected]; aloewen@eonerc. The focus of this work is on the accurate estimation of the
rwth-aachen.de; [email protected]; amonti@eonerc. amplitude and phase angle of voltage and current, which
rwth-aachen.de; [email protected]).
A. Benigni is with the Department of Electrical Engineering, University enables the accurate estimation of the power flows, possibly
of South Carolina, Columbia, SC 29208 USA (e-mail: benignia@ even under dynamic conditions.
mailbox.sc.edu). These types of state estimation approaches are becoming
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. increasingly sophisticated and yield highly valuable
Digital Object Identifier 10.1109/TIM.2015.2398991 knowledge. However, they require the availability of a system
0018-9456 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1301

model, a significant number of measurements [9]–[12], for the estimation of voltage magnitudes. It makes no use
and a significant computational effort. Consequently, the of the typical state estimation equations and is flexible with
application of WLS state estimation to distribution networks regard to the number of input measurements. Depending on
faces a considerable cost barrier and, hence, wide deployment the available measurements, the estimation accuracy may vary,
of these techniques is not to be expected in the near future. but there are no convergence issues. Furthermore, rather
Instead, if less information on the status of the grid is than performing centralized voltage estimation for the entire
sufficient, and less accuracy is acceptable, at least in the system, this approach proposes a hierarchical bottom-up archi-
current transition phase toward future distribution system tecture to build up the overall picture. Moreover, the estimation
automation, it is possible to use alternative approaches. This is process is computationally simple and can be executed on a
the case, for example, if the DSO is mainly interested in esti- low-cost hardware. Ferdowsi et al. [16] provided a demon-
mating only the voltage magnitude of the buses in the system, stration by implementing a prototype for embedded execution
to detect possible violation of permissible voltage magnitude using a single-board computer (BeagleBone Black) and testing
limits, i.e., undervoltages and overvoltages to comply with it in real time using data from a simulated LV feeder. In this
power quality standards, such as EN-50160 or IEC 61000, paper, we extend the work in [16] by introducing a new
for which the phase angle is not relevant; or if the control stage in the state estimation process that handles changes in
countermeasures, in response to violations detected by the the network configuration, which are common in MV level.
monitoring system, have fairly coarse granularity. In these Furthermore, we discuss the scalability of the proposed
cases, the accuracy requirements are significantly reduced. approach and present new test cases for configuration identifi-
Besides, the features of the estimation that we propose may cation (CI) and voltage magnitude estimation in the MV grid.
fit the requirements of other new monitoring and control In Section II, the monitoring approach is described.
services, which are still in development for the active Scalability aspects are explained and considerations presented
distribution networks of the future. for the design of local estimator and the ANN-based voltage
The use of an artificial neural network (ANN) for state estimator, CI unit, and data acquisition, transmission, and
estimation is a viable alternative to the classical model-based visualization. Section III introduces the prototype and
state estimation. An ANN is a data-driven approach and, the laboratory setups. Next, in Section IV, the approach is
therefore, it does not need any physics-based model of the demonstrated on two test case systems. Finally, the conclusion
system. It has an extremely light computational cost, making it is presented in Section V.
executable on very low cost hardware, and it requires few mea-
surements (with loose requirements in terms of accuracy and II. A PPROACH
synchronization). Such ANNs, when trained, accept substation The underlying idea in our approach is to have a scalable
measurements as inputs and provide node voltage magnitudes and low-cost solution that can unlock the door to the
as outputs. The knowledge of the behavior of the distribution development of monitoring and automation in distribution
grid is captured by the ANN coefficients and it is extracted, systems imposed by the cost barriers associated with the
during the training, from known load profiles and related implementation of classical approaches. Rather than providing
voltage profiles. Such data can typically be taken from load a very detailed picture and accurate estimation, our approach
flow solutions produced by the DSO (e.g., when assessing enables the operator to have a rough understanding of the
hosting capacity in response to a request to connect new ongoing conditions and, above all, to determine if there is a
generation or load), although other sources of information can voltage violation at all, and in the second instance to detect
equally easily be exploited. Note that training the ANN with the weakest points of the network without heavy investments.
selected data, such as working days only or peak hours, allows In this respect, our solution can constitute an intermediate
different sets of coefficients of the ANN to be determined and but affordable step in the long-term process of automation
appropriately used in the different day/time or other situations. of distribution systems. The follow up is envisioned as
Each such ANN computes the output for the portion of the incremental deployment of these local estimators until
grid connected downstream to the substation. Even a sparse the full distribution grid is monitored, integration of new
deployment provides insight into the status of some areas of measurements and full integration of the standalone Hardware
the network, although, in this case, partial but supportive of that embeds the estimator into the future standard substation
substation-level decision making. Hence, what we propose can automation architecture. Estimation of the voltage phase
be a cornerstone of the new monitoring and automation system angles is intentionally excluded from the monitoring problem
for distribution. as it is currently of little practical importance for DSOs in most
In [11], a state estimation technique based on ANNs for cases. The information provided by such a monitoring system
distribution systems is proposed. This technique, however, is beneficial both for the utility and the consumers: the utility
does not provide a modular and scalable structure. In some can use the information not only to better identify operational
cases, ANNs have been used in conjunction with WLS esti- limits of the network and take necessary corrective actions but
mation techniques, as in [13]–[15], where the ANN is used for also to make better decisions about network reinforcement.
the generation of the pseudomeasurements, thus constituting All this is expected to improve power quality for consumers.
a totally different application from what is proposed here. The bottom-up approach is realized by splitting the distrib-
Ferdowsi et al. [16] presented the initial version of this ution system into smaller parts, herein referred to as sections,
data-driven bottom-up monitoring approach based on an ANN and estimating node voltage magnitude within each section via
1302 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

Fig. 1. Scheme of the bottom-up hierarchical data-driven distribution monitoring approach.

an ANN. The idea is then to build the overall system picture power injection along the feeder from DG units could lead
by integrating the individual estimation solutions. The ANN, to very different voltage profiles, which all result in almost
as mentioned above, is trained with the data already available identical voltage and current measurements at the substation.
to the DSO from previous load flow calculations and by any In such cases, it is necessary to provide the estimator with
additional real-time measurements that may be available, from additional voltage or current measurements to capture the
smart meters, for example. This training results in an ANN that major voltage magnitude deviations. However, for most simple
mirrors the distribution system in terms of relation between network configurations, additional measurements are usually
voltage magnitude profile and load profile. In reality, though, not necessary. On the other hand, the estimator has the
measurements are affected by uncertainty. To incorporate this possibility to incorporate new measurements (e.g., smart meter
aspect, that is to link blurry input data to the same output, the data), when available, to improve its estimation accuracy by
ANN is trained with inputs corrupted by random noise. changing only the ANN code and without any need to change
The first step in splitting the system is to consider each its hardware implementation or the architecture of the mon-
portion of the LV grid downstream of an MV/LV transformer itoring system. Therefore, the estimation accuracy may vary
as a separate section for estimation and monitoring. In the depending on the number, location, and accuracy of available
event that different LV sections could be connected via measurements and depending on the complexity of the section
a reconfiguration switch, these sections must together be topology. Finally, in presence of DG units, the inclusion of
assigned one single local estimator. Similarly, at MV level, weather data (solar irradiation, wind speed, and so on) among
a portion of the system that exchanges power with other the inputs of the ANN may enable it to differentiate among a
portions of the grid only via transformers could be considered variety of different scenarios that can give rise to similar volt-
as a separate section, meaning that the MV feeders that could age and current measurements. Therefore, by having access
be connected via reconfiguration switches, as shown in Fig. 1, to weather data, the ANN may perform a more or equally
should be collectively considered as a single section. Next, accurate estimation with less demanding requirements both in
each local estimator, either at LV or MV level, estimates the terms of number and accuracy of measurements.
voltage magnitude of all nodes in its assigned portion of the The estimate of the voltage magnitudes of the downstream
grid using its ANN-based estimation unit. nodes makes it possible for the LV local estimators to detect
As a general rule, each local estimator needs voltage voltage-limit violations in their sections and issue necessary
and current measurements at the secondary side of the alarms. These data can be made available locally and be also
transformer(s) feeding the section. Depending on the com- shared periodically, e.g., on a 1 min basis, with the data
plexity of the section topology and whether or not significant collector. The data collector should preferably be located in the
distributed generation (DG) units are present, one or more DSO control center: this enables it to monitor all LV sections
additional measurements may be needed to ensure reasonable that are equipped with a local estimator. In a similar way,
estimation accuracy. The local estimator in the implementa- at the MV level, MV local estimators use the current and
tion presented in this paper is predominantly designed for voltage measurements at MV/LV transformers along with the
performing estimations in passive networks. Nevertheless, this measurements available directly at the MV level to estimate
estimator is still able to work reasonably well with relatively the voltage profile of the MV grid. Therefore, in a bottom-up
low penetration of DG units. To illustrate this, in Section V, way, it is possible to reconstruct the overall picture of the
we present the performance of the local estimator designed for distribution system. This concept is shown in Fig. 1. It should
a passive LV system in the presence of DG units. In [17], it is be noted that the estimations in the LV sections are not
shown that the presence of long side branches or considerable used as input to the estimation at the MV level because the
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1303

which was used for the training purpose in this paper. The
number of inputs to the ANN is equal to the number of
measurements collected from the grid, e.g., two and six for
the LV and MV grids in our test cases presented in Section IV.
The number of ANN outputs, i.e., output neurons, equals the
number of nodes for which the voltage magnitude is to be
estimated (9 and 14 nodes for our test LV and MV grids,
respectively). In addition, as explained in detailed in
Section IV-C, the hidden layer contains two neurons. For
the hidden and output layers, sigmoid and linear activation
functions were considered, respectively.
To train the ANNs, either archival load flow results, which
the DSOs usually already possess from network planning
studies, or a set of new power flows, obtained from the
Fig. 2. Block diagram of the local estimator execution.
network model, could be used. The elements which need to
be extracted from the load flow data are a set of input and
measurements at the MV/LV transformer, i.e., the measured output pairs that form the required data samples. The inputs
current and voltage values, fully represent the downstream are the same quantities that the trained ANN will use to
network from a circuit analysis point of view to the MV level. perform estimations and the outputs are the quantities that are
expected to be estimated. Thus, the ANN inputs are the voltage
and current magnitudes at the substation and, if needed, the
A. Architecture of the Local Estimator additional current and voltage magnitudes from other branches
Fig. 2 shows the block diagram of the local estimator. and nodes. Similarly, the voltage magnitudes of all nodes of
The main functions are as follows: interest are also extracted and used as the ANN output in the
1) acquisition of the measurement inputs; training process.
2) solution of the ANN-based estimation problem; The number of collected samples must be large enough
3) transmission of the measured and estimated values to to cover the whole span of possible load combinations and
the upper level and to the data collector, respectively. result in a comprehensive training data set. Therefore, a higher
The block actual configuration is needed only when the number of samples may be required as the network complexity
section may be reconfigured, which is usually the case only increases in terms of topology and the type of connected
at MV level. It should be stressed that this step is merely elements (the profile of loads, the presence of DG units,
responsible for the selection of appropriate ANN weights and so on). The training of the LV ANN must comprise
corresponding to the actual configuration of the section based a wide range of values for the load demand of each node
on the output of the so-called CI unit, which will be described (e.g., between 0% and 100% of the peak power at any given
in Section II-C, and does not deal with a physical model time), and, in fact, as each node represents only one or
of the section. The date and time block is used if different a few customers, the variability of operating conditions is
ANNs are trained for different operating points, as described fairly broad. For the MV grid, however, the load demand
in the previous sections. Therefore, if all loads in a section of each node, assumed to be centered around an expected
are of the same type, for example, if they are all residential, a value known from the load profile of that node, presents a
single ANN can usually adequately perform the estimations at relatively small deviation (e.g., within a ±10% of the expected
different times and no time and date input would be needed. value) for any given time interval, since each node usually
The input measurements can be also heterogeneous in type represents an aggregate of several loads. Another important
(e.g., voltage, current, active and reactive powers, as well as consideration is that if a section can be operated under
meteorological data like temperature, irradiation, and wind different configurations (such as the MV system presented in
speed). Section IV), its voltage profile would follow different patterns
and, as a result, dedicated ANNs must be adopted for each
configuration. Furthermore, if all the loads in a given section
B. ANN Structure and Training for Voltage Estimation have similar load profiles (e.g., all are residential as is the case
The training procedure tunes the weights and biases of the for the LV system presented in Section IV), only one single
ANNs to capture the relation between loads and node voltages ANN for different hours of the day may be sufficient as it can
in a compact form (in the form of a few weights and biases). be assumed that demand of all nodes undergo similar changes
In [18], the main architectures, fundamental equations, and the during the day. This is, however, not the case if different
most common learning techniques for ANNs are presented. nodes have different load profiles that follow various patterns
In our application, a feedforward ANN with one hidden layer with time.
was chosen for both the LV and MV local estimators. The Once trained offline, the ANNs are computationally
Levenberg–Marquardt algorithm was used to train the ANN. simple and may be implemented and deployed on a low-cost
This algorithm is a fast and efficient second-order method [18] hardware. There is also the possibility for retraining the ANN
and it is available in the MATLAB Neural Network Toolbox, upon availability of new data, which could be done remotely.
1304 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

Fig. 3. Diagram summarizing different modes of operation of the CI unit, the events triggering transition from one mode to another, and the choice of ANNs
in each mode.

It should be pointed out that the location and type of available the connectivity changes into the distribution management
measurements must be known for the initial ANN training. system (DMS) network model [20]. As a result, even in such
If some measurements are suddenly unavailable in runtime, cases, the actual network topology is available via the DMS.
the estimator will have to switch to an ANN that requires If the information about the current network configuration
fewer input measurements if possible, or issue an alarm is not available or accessible either directly from the FLIR
indicating that the estimator is missing some essential input system or the DMS at the control center, the CI unit would
data, and therefore, its output is not reliable. In practice, the have to detect the configuration itself. To this end, as the first
implementation on a low-cost device and the possibility to step, the CI must differentiate among three operating modes, as
initially avoid further investments in measurement devices shown in Fig. 3. The first mode is the normal operating mode
could make this solution economic and attractive for utilities. in which all reconfiguration switches, as shown in Fig. 1, are
For the implementation of the ANN-based estimation unit, open and feeders are fed in a radial way; the second mode
a dedicated solver was used. Since the training is performed corresponds to grid conditions in which part of the network
offline, the estimation is characterized by extremely low is deenergized as a result of a system fault and opening of
computational cost, which allows very cheap and relatively circuit breakers. In the third mode, which usually follows
slow processors (<1 GHz) to be used. the second mode, the faulty section of the system is isolated
by means of load break switches and an alternative path is
provided to supply the disconnected loads by closing one of
C. Configuration Identification the reconfiguration switches.
Unlike LV distribution systems, reconfiguration is common To detect the mode in which the grid is currently operating,
at MV level. More precisely, radially operated ring or meshed the CI unit relies on the status information collected from
networks allow for feeding the loads through alternative paths the reconfiguration switches, the current measurements from
in the event of permanent faults. Considering the possibility feeders, which are compared with the expected values from
of various fault locations and corresponding reconfigurations, historical data, and the detection of the sudden changes
it is important to ensure that MV local estimators can maintain in the feeder loading level. The normal operation mode is
their estimation accuracy in reconfigurations. characterized by all reconfiguration switches being open and
The appropriate CI approach for a given distribution section the feeder currents having no unusual deviation from historical
depends strongly on its level of automation. In the presence of data known at each time interval. Transition to mode 2 can be
advanced outage management systems, such as fault location, detected by a sudden drop in the load of a feeder in which
isolation, and service restoration (FLIR), the information a fault occurs, as a result of operation of circuit breakers.
from protective relays or fault passage indicators provides In this mode, the loading level of other feeders does not
the opportunity to locate and isolate the faults. Based on this undergo a similar change and the reconfiguration switches
information, the operator or the automated FLIR function are still open. Transition to mode 3 is detectable from the
executes the required switching steps to restore the loads in sudden increase in the overall MV section load back to the
the sections not affected by the fault [19], [20]. Therefore, the expected range for that time interval as a result of the provision
MV local estimator is able to directly update its selection of of an alternative supply path to part of the feeders with
the MV system configuration according to the actual network an isolated line segment through one of the reconfiguration
configuration as reflected in the FLIR and implemented by switches. During the transition from mode 3 back to mode 1,
the operator in the control center or the automated FLIR a simultaneous step increase in the loading of one feeder and
function. In practice, even in distribution systems with trouble a load decrease in another feeder is observed followed by the
call-based outage management system, the operator enters opening of reconfiguration switches.
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1305

If the grid is operating in mode 2 or 3, the CI unit

will then proceed with identifying the deenergized part
(which could consist of several branches and nodes) or
the isolated line segment of the section to determine the
actual configuration. For this task, the CI unit is equipped
with a library of ANN-based classifiers trained for different
hours of the day and months of the year. Training of these
ANN-based classifiers relies on the assumption that each load
on the MV usually represents a group of customers whose
aggregated demand follows standard load profiles (which vary
for different times of the year as well as weekdays and
weekends).
The CI unit also benefits from the current and voltage
measurements available from the LV estimation units in deter-
mining the actual mode of operation and identifying which part
of the section may be isolated. In this sense, for example, zero
current and voltage at the secondary of MV/LV transformers
could be a potential indicator of isolation of the MV node
from which the MV/LV transformer is fed. Furthermore, if the Fig. 4. Laboratory setup.
secondary substations have transformers with off-load tap
changers, whose tap settings are usually adjusted on a seasonal several European projects including FI for smart energy [21]
basis, the voltage measurement on the LV side can be used and FI smart utility services [22]. To this end, the central
to directly calculate the voltage magnitude on the MV side. cloud service is realized as a web service on a regular web
Such information can serve as additional information for the server where the data is stored in a database. This approach is
MV local estimator and its CI unit and significantly simplify expected to reduce the amount of data transfer as well as the
their task, if not completely remove the need for them, and connection time of the individual measurement or estimator
improve the accuracy of their estimations. devices.
It is also worth noting that the CI function described above For data collection and visualization, a web-based ad hoc
could be mapped to the topology process function in classical solution is developed. This visualization is useful not only
state estimation. for development and field test, but also for maintenance, as
it is independent of any existing visualization system. As a
D. Data Acquisition, Transmission, and Visualization
consequence, the field tests could be set up and carried out in
For measurement acquisition, the local estimator must a nonintrusive manner. The visualized measurements and esti-
be able to utilize a variety of technologies: in compatible mations are available through the web-based interface locally
substations, the IEC 61850 protocol could be used to acquire at the substation for the section where the local estimator is
sampled values signals over the process bus or the input working on, as well as globally in the control center for an
could be acquired from other intelligent electronic device overall picture. In this respect, the local visualization shows a
over the substation bus. In substations not compatible to the chart with real-time measurements used by the estimator and
IEC 61850 standards, analog signals can be used for locally the resulting estimations. The global visualization shows the
connected sensors, while remote measurements may come scheme of the grid model and the estimations or measurements
in via power line communication or wireless technologies. where available for each node. The voltage level is
To create a flexible and low cost device, the local estimator indicated by color coding ranging from green (normal) to red
hardware core can be built independently from the specific (limit exceeded). If the permissible voltage limits are violated,
communication media as a computational unit that offers alarms will be issued.
basic interfaces like USB for the connection of a wireless In parallel to the described web-based visualization,
communication device or an Ethernet port for communication an interface could be implemented to transmit data to a
with a local network or gateway. For the transmission of supervisory control and data acquisition (SCADA) system
the estimated values, a modular approach similar to the one using protocols like IEC 61850 or DNP3. The connection to a
applied for the measurements acquisition can be used. Further SCADA system would be also available locally and globally
details can be found in [16, Sec. II-B]. from the cloud service.
Even though the logical flow of information suggests direct
communication of the computed values to upper levels, it may
E. Scalability Aspects of the Proposed Approach
be more practical for each device to contact a cloud service
to deliver measurement data to the data collector and to The proposed monitoring approach is scalable in terms of
fetch data from the lower level measurement devices. Such the following aspects.
an approach would be also in line with recent trends in the 1) Modularity: The overall monitoring system is composed
application of future Internet (FI) concepts for smart energy of a number of estimation modules, namely, local
systems, which have been investigated and are being trialed in estimators, communicating with a central data collector.
1306 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

Each local estimator is responsible for a separate III. P ROTOTYPE AND L ABORATORY S ETUP
section of the system and acts independently from A. Prototype
other modules located at the same level. In this respect,
The prototype of the local estimator was implemented on a
transformers act as decoupling points between different
BeagleBone Black, which is a single-board computer equipped
sections. The number of monitored sections can be
with a 1-GHz ARM CPU and 512 MB of RAM. Among the
increased or reduced to monitor a larger or smaller
many similar boards available on the market, including its very
part of the overall distribution system by the addition
well-known competitor Raspberry Pi, the BeagleBone Black
or removal of estimation modules without any change
was chosen mainly for two reasons, namely, the presence of
in the overall architecture of the monitoring system.
an embedded storage of 2 GB and the capability of directly
Furthermore, as mentioned before, this approach has
acquiring analog signals. The embedded storage of 2 GB
the possibility for partial deployment and providing
together with a network booting system will considerably
monitoring in selected parts of the grids in the first
simplify future tests involving several local estimators. The
stages thanks to its modular architecture.
capability of directly acquiring analog signals is extremely
2) Extensibility: The same local estimator can be used
useful since input analog signals generated by a real-time
regardless of the number of the nodes and complexity of
digital simulator (RTDS) are used as measurements in the
the network of interest. The only difference may be the
tests. To send the estimated values to the data concentrator,
need for a higher number of measurement inputs in case
the prototype is equipped with a 3G stick (Huawei E3131).
of increased complexity in network topology, number of
The ANN-based estimation is performed using a custom-made
nodes, or the presence of significant DG. As described
solver executed on the BeagleBone Black.
in [16], it is also possible to design the local estimator
to seamlessly switch from its normal ANN to a
more basic ANN if one of the input measurements B. Laboratory Setup
is suddenly disconnected. In such a case, the local Fig. 4 describes the laboratory setup realized to test the
estimator must issue an alarm indicating that the proposed approach. The setup comprises RTDS, a commercial
accuracy of the estimations may be degraded compared tool for real-time simulation of power systems, the local
with the normal case. Furthermore, this solution is estimator prototype, and a web server implemented on a
extensible in the sense that its estimators can integrate normal PC. The distribution network model is implemented
new measurements upon availability. In such cases, only in RTDS. The prototype of the local estimator receives its
the ANN code must be updated in a remote way. Since measurement inputs through analog output board of RTDS.
computation of the ANN outputs is very fast (around The estimate is then sent out via 3G-based communication to
0.5–1 ms for the heaviest cases in our tests for the both the web server.
LV and MV systems; this is explained in more detail
in Section V-D), the same hardware could handle much IV. T EST C ASES
larger grids involving a higher number of ANN inputs
In this section, the capability of the proposed ANN-based
and outputs. In case the section can be reconfigured by
system to accurately estimate the voltage profile at different
means of reconfiguration switches, as shown in the MV
voltage levels of two typical European distribution systems is
level of Fig. 1, the local estimator should be equipped
investigated. The absolute value of the error in the voltage
with a CI unit, as discussed in detail in Section II-C.
magnitude estimation in the node with the largest error is
determined in each case. Furthermore, the ability of the ANNs
to detect voltage magnitude violations for each of the test
F. Applicability of the Proposed Approach systems is studied.
It should be emphasized that the proposed approach as
presented in this paper is applicable to networks with no or A. Test Distribution Systems
a relatively low penetration level of DG units. Furthermore, The tests are performed on one LV network and one MV
only radial or radially operated ring and mesh networks are distribution system shown in Fig. 5, which are obtained by
covered in this paper. Both voltage and current values at the slight modifications of the networks detailed in [23] and [25].
secondary side of the transformers also need to be measured The LV section represents a system supplying residential
and made available to the local estimators in all cases (further customers with a peak load demand of 258 kVA and a
measurements may be required depending on the topological load profile ranging between 20% and 100% of the peak
complexity). In Section V-A, the performance of the ANNs value during the day. The system is fed from a 300-kVA
trained for the operation in a test LV network with no DG units transformer and underground cables are used to distribute the
is presented when a DG unit is connected to the system. It is power to the loads. The MV section is derived from a German
shown that the estimation error increases with an increase in distribution network and is composed of two subnetworks,
the injected power from the DG unit. However, depending as detailed in [25]. Each subnetwork is supplied through
on the accuracy requirements of the application, the increased a transformer with a rated power of 15 MVA with most
estimation error may still be acceptable and the estimator may connections made via cables. The loads are a combination
be still used without adaptation for DG units. of residential and industrial loads. The peak demands of
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1307

Fig. 5. Topology of (a) test MV distribution system [25] and (b) LV residential feeder [23], [24].

TABLE I magnitudes, which are expected to be estimated for each of

S UMMARY OF I NPUTS AND O UTPUTS OF THE ANNs the LV and MV networks.
FOR LV AND MV L OCAL E STIMATORS In the ANN training phase, these measurement inputs form
the ANN input vector and the desired estimator outputs form
the ANN output vector.
For the LV network, 40 000 different samples were
generated. The load demand of each node in each sample was
considered to be a random number with a uniform distribution
from 0 to the peak demand of that node, as detailed in [16]. For
the MV system, considering different possible configurations,
residential and industrial loads in subnetwork 1 equal 58 and dedicated ANNs were trained for each envisaged case.
28 MVA, respectively, while in subnetwork 2, the loads have For each configuration, 40 000 samples were generated to
peak values of 4.5 and 23 MVA, respectively. The ranges train the ANN. The deviation of the load demand from the
of the load profile for residential and industrial loads in the expected value at each time interval was assumed to be ±10%.
MV system are from 20%–100% and 30%–100%, respectively, To tackle the measurement uncertainties faced during the
as presented in [25]. hardware implementation, random noise with a normal dis-
tribution was, as mentioned in Section II, added to each
input measurement of each generated sample. The accuracy
B. Training Data for ANN-Based Estimators
of voltage measurements was assumed to be 1% of the full
For the above LV section, it was shown in [16] that two scale of the measurement device, which is considered to
measurements from the section, i.e., the voltage and current be equal to the maximum value of the voltage. Therefore,
values at the LV side of the MV/LV transformer are adequate the added noise had an expected value of 0 and a standard
to perform a sufficiently accurate estimation of the voltage deviation equal to 0.33% of the maximum value of the voltage.
magnitude at the nodes of interest. This is due to the relatively For the current measurement, the accuracy class of 1% was
simple topology of the system that does not have long side assumed and it was assumed that this accuracy depends on the
braches and the lack of a significant share of the installed DG. reading value. More precisely, the current (ratio) errors were
For the MV section, the voltage and current values at the assumed to be 1%, 1.5%, and 3% of the reading value for
MV side of the feeder transformers as well as the current the current measurements at 100%, 20%, and 5% of the full
measurement in line L 3−4 and line L 3−8 are the required scale of the current transformer, respectively [26]. As a result,
inputs for the voltage profile estimation in the MV network depending on the value of the current in each sample, noise
assuming that the actual network configuration is known. with appropriate standard deviation was added to it. This was,
Table I summarizes these measurement inputs and the voltage however, an example, and in the general case, the uncertainty
1308 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

model and its associated parameters including the standard TABLE II

deviation could be modified based on the expected uncertainty R ECONFIGURATION S WITCHES , L INES S EGMENTS C OVERED BY E ACH OF
in the field implementation [27]. T HEM , AND THE C ODE C ORRESPONDING TO E ACH L INE

C. ANN Selection and Training

The ANNs with feedforward structure were selected and
trained with different numbers of neurons in the hidden
layer. It was observed that the best results in terms of the
ANN estimation accuracy and generalization capability were
achieved when 2 neurons were used in the hidden layer for where N is the total number of target values in the data set
both LV and MV grids. Inclusion of more neurons led to (i.e., total number of target voltage magnitudes for all nodes
overfitting of the ANN. However, in general, a higher number and for all data samples) and yi and z i are the values of the
of neurons may be needed in the hidden layer to ensure ANN estimates and the corresponding target values (for each
acceptable estimation accuracy. This might be the case, for node and each data sample), respectively [29].
example, when the number of required measurement inputs These values are reported in Sections V-A and V-C for each
for the ANN increases as a result of increased topological of the trained ANNs used for the LV and MV networks.
complexity of the network or significant local generation. D. Realization of the CI Unit
To figure out the optimal number of neurons in the hidden
layer, the general procedure is to consider two neurons in In mode 2, considering that no configuration switch is
the first place and only if the accuracy of the estimation closed and considering the radial operation of each feeder of
is not high enough and to increase the number of neurons the MV section, the amount of load drop detected by current
until a reasonable accuracy is achieved without compromising measurements provides a means for estimating the parts which
the ANN generalization capability. However, the number of are deenergized as a result of operation of a circuit breaker.
neurons should be kept as low as possible to maintain the In mode 3, the first step for realization of the CI unit
ANN generalization capabilities [28]. is to understand the possible grid configurations that the
The ANNs with the aforementioned ANN structure CI unit needs to differentiate between and classify. This means
were trained using the generated training data in the determining all possible configurations after the isolation of
MATLAB Neural Network Toolbox. The data samples the line segment due to the fault and closure of the reconfigu-
(40 000) were divided into training, validation, and test sets ration switch. For this purpose, various line outage scenarios
(50%, 25%, and 25%, respectively). The training set is used that require grid reconfiguration via closure of each of the
directly to tune the weights and biases of the ANN, while the reconfiguration switches are identified and listed, as shown
validation set is used to check the ANN during the training in Table II. For simplicity, a code has been assigned to the
phase to ensure that the generalization capability of the ANN line segments in each scenario.
is maintained. To assess the performance of the trained ANNs, As mentioned earlier, it is assumed that the CI unit
the test set (10 000 data samples), which was not used in the receives information regarding the status of the reconfiguration
training process and therefore could be regarded as a new data switches. This could be realized either through direct input
for the ANN, was used. Two performance indicators, namely, by the operator or by means of additional measurements
the regression value, i.e., the correlation coefficient (R), and monitoring the status of these switches, since the measure-
the mean square error (MSE), were calculated. The R values ments used for voltage magnitude estimation are not adequate
measure the correlation between the ANN estimates and the for identifying the closed reconfiguration switch. Knowing
output elements corresponding to different input elements in which reconfiguration switch is closed, the CI proceeds with
the data sample, namely, the target values (each data sample identifying the line code where the fault has occurred. This is
consists of an input element and its corresponding output). realized by using a dedicated ANN for each of the reconfig-
It can be shown that for a linear regression, the correlation uration switches. The ANN used for this purpose also has
coefficient (R) can be obtained from the following: a feedforward structure with the same six inputs as those
 used for the MV voltage estimation and five hidden neurons
ss 2yz and one output. The output of the ANN is a real number
R= . (1) that is then rounded to an integer value representing the line
ss yy ss zz
code as defined in Table II. In this example, since current
In this equation, ss yz is the unbiased estimation of the measurements are placed to directly measure the currents of
covariance of the ANN outputs and the corresponding target lines L 3−4 and L 3−8 , their isolation can be directly detected,
values, and ss yy and ss x x are the unbiased estimations of and therefore, there is no need to use the estimation process
the variances of the ANN outputs and the target values, to detect the isolated line. As a result, if S2 or S3 are closed
respectively [29]. The MSE can be also calculated from the and the isolation of L 3−4 and L 3−8 is not detected, the CI
following: only needs to differentiate among three the line segments.
Based on the above-mentioned considerations, the block
1 
N
MSE = (yi− z i )2 (2) diagram of the hardware implementation for the LV and MV
N local estimators will be as shown in Figs. 6 and 7.
i=1
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1309

Fig. 6. Block diagram of the hardware implementation of the LV local estimator.

Fig. 7. Block diagram of the hardware implementation of the MV local estimator.

Fig. 8. Comparison of measured and estimated values for the LV section with (a) light, (b) medium, and (c) heavy load scenarios.

V. R ESULTS
The LV and MV grids were simulated in RTDS and the
voltage of each node in the feeder was recorded. At the same
time, through the giga transceiver analog output cards of the
RTDS, the measurements were fed into the local estimator.

A. Voltage Estimation at LV Level

For the LV system, the trained ANN was first tested
with 10 000 sample data and its outputs (i.e., the estimated Fig. 9. Estimation error of the LV local estimator in the presence of power
node voltages) were compared with the target values injection from a DG unit installed at the end of the LV feeder.
(i.e., the measured node voltages). The MSE and correlation
coefficient (R) were found to be 0.5538 and 0.96767, which detailed in [16]. In Fig. 8, a comparison between the measured
indicate that the trained ANN can accurately perform voltage voltage magnitudes and the values estimated by the local
magnitude estimations. Then for the hardware test, three estimator are provided at different nodes. It was observed that
scenarios of light, medium, and heavy load with total loads of the estimation accuracy is very high (error ≈ 1%) for all the
83.6, 161.1, and 237.2 kVA, respectively, were considered, as cases. It should be also noted that the increase in the voltage
1310 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

Fig. 10. Comparison of the actual and estimated line codes before rounding for the case of reconfiguration by the closure of switches
(a) S1, (b) S2, and (c) S3.

magnitudes in the middle of these graphs does not indicate

an increase in the voltage along the feeder but because on
the x-axis of the figures, bus 15, located at the end of a side
branch, is placed before bus 6 on the main feeder. It should
be noted that the estimator has also properly estimated the
violation of the voltage magnitude in the third scenario.
The horizontal line shown in these figures indicates the
lower limit of the permissible voltage for the LV feeder.
According to the standard IEC 61000, under normal system
operating conditions, the voltage magnitude variations at a
network user’s supply terminal in the public LV and MV
distribution networks must be within 10% of the nominal
voltage. However, this variation range includes voltage drops Fig. 11. Comparison of the actual and estimated line code before rounding
for the case of reconfiguration by the closure of switch S1 when the load
or rises in the HV/MV transformer, MV network, MV/LV in node 13 (connected to its neighboring nodes by lines represented by
transformer, and LV network collectively. Different operators codes 3 and 4) is very small.
may allocate this permissible range differently to each of these
four parts. In our tests, the allocation of voltage variation range TABLE III
is done based on [30], thus considering 7% voltage drop for P ERFORMANCE OF THE T RAINED ANNs IN E STIMATION OF 10 000 T EST
each of the MV and LV networks together with their feeding D ATA FOR E ACH S AMPLE S CENARIO
transformers (assuming that the voltage at the HV side of the
HV/MV substation is about 3% higher than the nominal value
to allow for higher voltage drop at MV and LV levels). It is
also assumed that the voltage at the MV side of the MV/LV
transformer is 19.4 kV for all the three scenarios as a result
of 3% voltage drop in the MV network. Thus, the voltage at
the LV side of the MV/LV transformers would be 223.1 V if
the transformer would not cause any voltage drop. However,
due to the presence of transformer equivalent impedance, it can
be observed from Fig. 8 that the voltage at the beginning
of the LV feeder decreases as it becomes more heavily
loaded. estimation error in the system as the DG power injection
To assess the performance of the above estimator, which was increases from 0 to 80 kW is shown. On the x-axis, the injected
trained to estimate in a purely passive network, in the presence power from the DG unit is also specified as a percentage of
of DG units, a test case was studied: a DG unit was assumed the load at node 18, which was assumed to be 22.04 kW in
to be installed at the end of the LV feeder, i.e., node 18 and the light-load scenario [16].
the light load scenario as described in the previous example As shown in the figure, the estimation error increases with
was considered. This combination of DG location and loading the increase in the injected power from the DG unit. Although
scenario was selected to present a case in which the power the increased estimation error is not a desirable outcome,
injection from the DG unit has a relatively high impact on a certain level of estimation error may be still acceptable
the voltage profile [31] and therefore causes possibly higher depending on the application. Thus, the estimator with the
estimation errors. The DG unit was assumed to inject only same ANN may be still usable in the presence of a certain level
active power. In Fig. 9, the absolute value of the largest of power injection DG units depending on the application.
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1311

Fig. 12. Comparison of the measured and estimated values for the MV level for (a) normal operation with all the three configuration switches open,
(b) when line L 1−2 is disconnected and S1 is closed, (c) when line L 8−9 is disconnected and S3 is closed, (d) when line L 13−14 is isolated and S1 is closed,
(e) when line L 4−5 is isolated and S2 is closed, and (f) when the transformer of subnetwork 1 is isolated and S1 is closed.

B. Configuration Identification demand at nodes 12 and 14. It is also assumed that the load
demand at each node at any given time interval is centered
Fig. 10 presents the raw output of the ANN used in the around an expected value with a ±10% deviation. As a result,
CI unit before rounding for a number of different loading the current measurements at substation of subnetwork 2 and
scenarios and for all the thee cases of reconfiguration realized line L 3−8 cannot clearly indicate if node 13 is supplied
by the closure of S1, S2, or S3. As shown in the figure, the through substation 2 or substation 1 (via S1). Therefore, the
isolated line is correctly estimated in all the three cases. CI unit may not accurately determine which line segment in
It is possible that instead of the isolated line, one of the subnetwork 2, namely, L 12−13 or L 13−14 , is isolated. The
lines next to it is identified by the CI as the isolated one. estimated line code by the CI unit before rounding in such
In general, this could happen when the loading level of one a scenario is shown in Fig. 11 for 100 tests corresponding to
or more of the loads, whose total current is measured by each line code.
a single current transducer, is considerably smaller than the The consequence would be that a slightly different configu-
others. In such case, these nodes have very little impact on ration is considered in which the estimated voltage magnitude
the overall measured current. As a result, considering that the of one node may deviate significantly from its actual value,
actual loading of all nodes is uncertain, using the sum of the because it is considered to be fed from a path other than the
currents of the nodes, it is not possible to accurately guess if real one. Such errors could be detected and fixed manually
the nodes with relatively small currents are contributing to the by comparing the actual configuration with the one estimated
measured current value or not. and used by the CI that is available to the operator through
To illustrate this, consider the case that a fault has occurred the visualization system explained in Section II-D.
in subnetwork 2, the faulty line segment has been isolated,
and S1 is closed to provide an alternative path to feed the
loads that are not fed through substation 2 anymore. Let us C. Voltage Estimation at MV Level
assume that nodes 12, 13, and 14 have a similar load profile, For the MV system, six scenarios were considered for a
but the peak demand at node 13 is only 10% of the peak fixed loading condition in which subnetworks 1 and 2 had
1312 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 5, MAY 2015

total loadings of 6.9 MVA and 6.2 MVA (about 80% and VI. C ONCLUSION
90% of their peak loads), respectively. In the first scenario, This paper presents a novel solution for monitoring
the system was in a normal operating mode with all the distribution systems based on a data-driven bottom-up scalable
reconfiguration switches open. In the second scenario, it was approach that uses ANNs for its estimations. Using experi-
assumed that line L 1−2 had been disconnected due to a fault mental tests, it was shown that ANN-based local estimators,
and switch S1 was closed, allowing part of the loads on feeder which are the main building blocks of the proposed monitoring
1 to be supplied through feeder 2. In scenario 3, line L 8−9 was system, are capable of estimating the voltage profiles of
assumed to have been disconnected and switch S3 was closed LV and MV sections of a distribution system when only a
to provide an alternative path for the loads originally supplied few measurements are available. To improve the accuracy of
via line L 8−9 . Similarly, in the fourth and the fifth scenarios, the estimation in the presence of measurement uncertainties,
line segments L 13−14 and L 4−5 were assumed to be sec- the ANNs are trained with artificially degraded data. In the
tionalized and switches S1 and S2 were closed, respectively. presence of reconfiguration switches, an add-on CI unit detects
The sixth scenario illustrates a case where the transformer the actual configuration of the system and loads the cor-
of subnetwork 1 was disconnected and S1 was closed to responding ANN in the local estimator. The realization of
provide an alternative path for all the loads in subnetwork 1. the CI unit requires the determination of possible scenarios
The ANNs used for each of these scenarios were first of grid configuration in advance and the use of a classifier
checked using 10 000 test data in MATLAB. The correlation to differentiate among them in the real-time operation.
coefficient (R) and the MSE values, which compare the ANN Various scenarios of CI implementation are described and a
outputs and the targets values, are provided in Table III. web-based visualization approach for the monitoring system is
As Table III shows, the trained ANNs are capable of introduced. The modularity, extensibility, requirement of only
estimating the node voltage magnitudes with high accuracy. few measurements to provide valuable insight into system
Next, the ANN were implemented on the laboratory setup operating status, and the possibility for low-cost hardware
and tested for one arbitrary operating point in each of these implementation makes this approach stand out as a very practi-
six scenarios. The actual values of the voltage magnitudes cal solution for DSOs to monitor their systems at a reasonable
were compared with the values estimated by the MV local cost compared with other proposed solutions such as PMUs.
estimator and the lower limit of the voltage in Fig. 12. It is
R EFERENCES
can be observed that the estimations are highly accurate (with
[1] EU Commission Task Force for Smart Grids Expert Group 3.
the maximum error not exceeding about 1% in most cases). (Apr. 2011). Roles and Responsibilities of Actors Involved in the
Furthermore, violation of voltage limits in the last scenario is Smart Grids Deployment. [Online]. Available: http://ec.europa.eu/
detected properly in scenario 6. energy/gas_electricity/smartgrids/doc/expert_group3.pdf
[2] Terna. [Online]. Available: http://www.terna.it, accessed Feb. 19, 2015.
[3] ENEL Annual Report for 2012. [Online]. Available:
D. Execution Time http://annualreport2012.enel.com, accessed Feb. 19, 2015.
[4] R. Singh, B. C. Pal, R. A. Jabr, and R. B. Vinter, “Meter placement for
Although the presented ANN-based approach is not distribution system state estimation: An ordinal optimization approach,”
directly comparable with WLS method, in the sense that it IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2328–2335, Nov. 2011.
only estimates the voltage magnitude rather that the full state [5] M. Pau, P. A. Pegoraro, and S. Sulis, “Efficient branch-current-based dis-
tribution system state estimation including synchronized measurements,”
of the system and requires only very few measurements, IEEE Trans. Instrum. Meas., vol. 62, no. 9, pp. 2419–2429, Sep. 2013.
it would be insightful to provide a rough comparison of the [6] J. Liu, J. Tang, F. Ponci, A. Monti, C. Muscas, and P. A. Pegoraro,
two techniques here. In our tests, the execution time of the “Trade-offs in PMU deployment for state estimation in active distri-
bution grids,” IEEE Trans. Smart Grid, vol. 3, no. 2, pp. 915–924,
ANN-based estimators was observed to be around 0.5–1 ms Jun. 2012.
for both LV and MV test systems. In [5], the execution time of [7] C. Muscas, M. Pau, P. A. Pegoraro, and S. Sulis, “Effects of measure-
WLS method for an 18-bus system, which is exactly the same ments and pseudomeasurements correlation in distribution system state
estimation,” IEEE Trans. Instrum. Meas., vol. 63, no. 12, pp. 2813–2823,
size of the LV system used in our test case, has been reported Dec. 2014.
to be between 2.0 and 6.7 ms depending on the state model [8] P. Castello, J. Liu, C. Muscas, P. A. Pegoraro, F. Ponci, and A. Monti,
(polar or rectangular) and the measurements (type 1 or 2). “A fast and accurate PMU algorithm for P+M class measurement of
synchrophasor and frequency,” IEEE Trans. Instrum. Meas., vol. 63,
The hardware used in this reference is an Intel Dual Core no. 12, pp. 2837–2845, Dec. 2014.
3.17-GHz CPU (Core 2 Duo E8500, >30 000 DMIPS), which [9] I. Roytelman and S. M. Shahidehpour, “State estimation for electric
is at least 15 times faster than the 1-GHz ARM Cortex-A8 power distribution systems in quasi real-time conditions,” IEEE Trans.
Power Del., vol. 8, no. 4, pp. 2009–2015, Oct. 1993.
CPU (2000 DMIPS) of Beaglebone Black. This means that if [10] M. E. Baran and A. W. Kelley, “State estimation for real-time moni-
the WLS method were run for the 18-bus system described toring of distribution systems,” IEEE Trans. Power Syst., vol. 9, no. 3,
pp. 1601–1609, Aug. 1994.
in [5] on the Beaglebone Black, it would take 15 times longer [11] A. Bernieri, G. Betta, C. Liguori, and A. Losi, “Neural networks and
to execute, which means 30 to 100.5 ms (assuming similar pseudo-measurements for real-time monitoring of distribution systems,”
measurements, this should be also the approximate time for IEEE Trans. Instrum. Meas., vol. 45, no. 2, pp. 645–650, Apr. 1996.
[12] C. Muscas, S. Sulis, A. Angioni, F. Ponci, and A. Monti, “Impact
running the WLS method for our LV system). of different uncertainty sources on a three-phase state estimator for
Although, according to the above comparison, the ANN- distribution networks,” IEEE Trans. Instrum. Meas., vol. 63, no. 9,
based method is considerably faster than the WLS technique, pp. 2200–2209, Sep. 2014.
[13] M. Biserica, Y. Besanger, R. Caire, O. Chilard, and P. Deschamps,
it should be stressed that the main advantage of the ANN- “Neural networks to improve distribution state estimation—Volt var
based method is that it requires only very few measurement control performances,” IEEE Trans. Smart Grid, vol. 3, no. 3,
orders to provide valuable insight into the system status. pp. 1137–1144, Sep. 2012.
FERDOWSI et al.: SCALABLE DATA-DRIVEN MONITORING APPROACH 1313

[14] J. Wu, Y. He, and N. Jenkins, “A robust state estimator for medium Andrea Benigni (S’09–M’14) received the
voltage distribution networks,” IEEE Trans. Power Syst., vol. 28, no. 2, B.Sc. and M.Sc. degrees from the Politecnico
pp. 1008–1016, May 2013. di Milano, Milan, Italy, in 2005 and 2008,
[15] E. Manitsas, R. Singh, B. C. Pal, and G. Strbac, “Distribution system respectively, and the Ph.D. degree from RWTH
state estimation using an artificial neural network approach for pseudo Aachen University, Aachen, Germany, in 2013.
measurement modeling,” IEEE Trans. Power Syst., vol. 27, no. 4, He was a Research Associate with the Institute
pp. 1888–1896, Nov. 2012. for Automation of Complex Power System, E.ON
[16] M. Ferdowsi, A. Benigni, A. Löwen, P. McKeever, A. Monti, and Energy Research Center, RWTH Aachen University,
F. Ponci, “New monitoring approach for distribution systems,” in from 2009 to 2013. He is currently an Assistant
Proc. IEEE Int. Instrum. Meas. Technol. Conf. (I2MTC), Montevideo, Professor with the Department of Electrical
Uruguay, May 2014, pp. 1506–1511. Engineering, University of South Carolina,
[17] M. Ferdowsi, B. Zargar, F. Ponci, and A. Monti, “Design consider- Columbia, SC, USA.
ations for artificial neural network-based estimators in monitoring of
distribution systems,” in Proc. IEEE Int. Workshop Appl. Meas. Power
Syst. (AMPS), Aachen, Germany, Sep. 2014, pp. 1–6.
[18] B. M. Wilamowski and J. D. Irwin, Intelligent Systems (Industrial
Electronics), 2nd ed. Boca Raton, FL, USA: CRC Press, 2011.
[19] R. A. A. de Graaff, “Flexible distribution systems through the appli- Artur Löwen completed his three years of training
cation of multi back-to-back converters: Concept, implementation and as a Computer Science Expert for Application
experimental verification,” Ph.D. dissertation, Dept. Electr. Energy, Development. He was with Stöhr+Sauer, Würselen,
Eindhoven Univ. Technol., Eindhoven, The Netherlands, 2010. Germany, where he was involved in database-
[20] J. Northcote-Green and R. G. Wilson, “Central control and powered CAD applications and 3-D visualizations,
management—Outage management,” in Control and Automation of before joining the Institute for Automation of
Electrical Power Distribution Systems. Boca Raton, FL, USA: Complex Power Systems, RWTH Aachen
CRC Press, 2007, pp. 50–60. University, Aachen, Germany, in 2013, where he is
[21] FINSENY Project. [Online]. Available: http://www.fi-ppp-finseny.eu, currently involved in multiagent systems, embedded
accessed Feb. 19, 2015. systems, data protocols, and data processing
[22] FINESCE Project. [Online]. Available: http://www.finesce.eu, accessed automation.
Feb. 19, 2015.
[23] S. Papathanassiou, N. Hatziargyriou, and K. Strunz, “A benchmark low
voltage microgrid network,” in Proc. CIGRE Symp., Athens, Greece,
Apr. 2005, pp. 1–8. Behzad Zargar received the B.Sc. degree in elec-
[24] M. Ferdowsi. (Dec. 2011). Power Quality Assessment MERGE (Mobile trical power engineering from the Islamic Azad
Energy Resources in Grids of Electricity) Project, Deliverable D2.3. University of Najafabad, Isfahan, Iran, in 2009. He is
[Online]. Available: http://www.ev-merge.eu currently pursuing the M.Sc. degree in electrical
[25] K. Rudion, A. Orths, Z. A. Styczynski, and K. Strunz, “Design of power engineering with RWTH Aachen University,
benchmark of medium voltage distribution network for investigation Aachen, Germany.
of DG integration,” in Proc. IEEE Power Eng. Soc. General Meeting, His current research interests include power sys-
Montreal, QC, Canada, Jun. 2006, pp. 1–6. tems monitoring, grid integration of renewable
[26] Instrument Transformers Ltd. Current Transformer Accuracy. energy, and control of power systems.
[Online]. Available: http://www.itl-uk.com/en/services/knowledge-
bridge/accuracy-class, accessed Feb. 19, 2015.
[27] A. Ferrero and S. Salicone, “Uncertainty: Only one mathematical
approach to its evaluation and expression?” IEEE Trans. Instrum. Meas.,
vol. 61, no. 8, pp. 2167–2178, Aug. 2012. Antonello Monti (M’94-–SM’02) received the
[28] D. Hunter, H. Yu, M. S. Pukish, J. Kolbusz, and B. M. Wilamowski, M.Sc. (summa cum laude) and Ph.D. degrees in elec-
“Selection of proper neural network sizes and architectures—A com- trical engineering from the Politecnico di Milano,
parative study,” IEEE Trans. Ind. Informat., vol. 8, no. 2, pp. 228–240, Milan, Italy, in 1989 and 1994, respectively.
May 2012. He started his career with Ansaldo Industria,
[29] N. R. Draper and H. Smith, Applied Regression Analysis (Wiley Series Milan, and joined the Politecnico di Milano as an
in Probability and Statistics), 3rd ed. Hoboken, NJ, USA: Wiley, 1998. Assistant Professor in 1995. In 2000, he joined
[30] “Active distribution system management—A key tool for the smooth the Department of Electrical Engineering, University
integration of distributed generation,” in Proc. EURELECTRIC, of South Carolina, Columbia, SC, USA, as an
Feb. 2013. Associate Professor and then a Full Professor. Since
[31] H. M. Ayres, W. Freitas, M. C. de Almeida, and L. C. P. da Silva, 2008, he has been the Director of the Institute for
“Method for determining the maximum allowable penetration level Automation of Complex Power System with the E.ON Energy Research
of distributed generation without steady-state voltage violations,” IET Center, RWTH Aachen University, Aachen, Germany. He has authored or
Generat., Transmiss. Distrib., vol. 4, no. 4, pp. 495–508, Apr. 2010. co-authored over 200 peer-reviewed papers in international journals and
proceedings of international conferences.
Dr. Monti is an Associate Editor of the IEEE S YSTEM J OURNAL.

Mohsen Ferdowsi (S’14) was born in Esfahan, Iran,

in 1984. He received the B.Sc. and M.Sc. degrees in
electrical engineering from the University of Tehran, Ferdinanda Ponci (M’00–SM’08) received the
Tehran, Iran, in 2007 and 2009, respectively. He is M.S. and Ph.D. degrees in electrical engineering
currently pursuing the Ph.D. degree with RWTH from the Politecnico di Milano, Milan, Italy,
Aachen University, Aachen, Germany. in 1998 and 2002, respectively.
He was a Research Associate with the Technische She joined the Power and Energy Research Group,
Universität Berlin, Berlin, Germany, from 2009 Department of Electrical Engineering, University of
to 2011. Since 2012, he has been a Research South Carolina, Columbia, SC, USA, in 2003, as an
Associate with RWTH Aachen University. Assistant Professor, and was tenured and promoted
His current research interests include application of to Associate Professor in 2008. In 2009, she joined
data-driven approaches to power system monitoring, co-simulation of power the Institute for Automation of Complex Power
and communication systems, power quality, power system dynamics, and Systems, E.ON Research Center, RWTH Aachen
grid integration of renewable and distributed energy resources. University, Aachen, Germany, where she is currently an APL Professor.