A Neural Multi-agent-based Approach for Preventing Blackouts
in Power Systems
Michael Negnevitsky
1
, Nikita Tomin
2
, Daniil Panasetsky
2
, Ulf Haeger
3
, Nikolay Voropai
2
,
Christian Rehtanz
3
and Victor Kurbatsky
2
1
School of Engineering, University of Tasmania, Churchill Ave, Sandy Bay TAS 7005, Hobart, Australia
2
Department of Electric Power Systems, Melentiev Energy Systems Institute, Lermontov str. 130, Irkutsk, Russia
3
Energy Systems, Energy Efciency and Energy Economics, TU Dortmund, Emil-Figge-Str. 70, Dortmund, Germany
Keywords: Blackout, Pre-emergency Control, Voltage Stability, Multi-agent Control System, Artificial Neural
Networks.
Abstract: A neural multi-agent-based approach for system monitoring and preventing large-scale emergencies in
power systems is presented in this paper. The automatic emergency control process is represented as a
neural multi-agent system with hierarchical architecture. The proposed system consist of two main parts: the
alarm trigger, a Kohonen neural network-based system for early detection of possible alarm states in a
power system, and the competitive–collaborative multi-agent control system. For demonstration purposes,
we investigated conventional and neural multi-agent automatic control schemes. Results are presented and
discussed.
1 INTRODUCTION
The ongoing deregulation and restructuring in power
system worldwide require more complex control and
decision making. In many cases, the current
generation of automatic emergency control systems
is ineffective and unreliable. Moreover, in
emergency condition, a power system operator has
to deal with a large amount of data and apply most
appropriate remedial actions. At such times it
becomes difficult to reach a correct diagnosis of the
problem or to formulate the correct decision when
actions must be taken. As a result, large scale
blackouts still happen (PSDP 2007, Wang 2005)
Computational intelligence techniques in power
systems provides a way forward to give new
possibilities for energy management systems,
especially in the field of preventing large scale
emergencies. There are many benefits to using a
multi-agent system as automatic control system,
such as the ability to perform multiple
computationally intensive tasks in parallel such that
effective optimized real-time control can be
achieved. These parallel tasks include neural
network training, parameter optimization, and
system monitoring. The multi-agent system
approach also allows for intelligent control that is
robust and flexible in that it can autonomously make
decisions and adjust to partial control system failure
to maintain control with minimal performance
degradation, to name a few of the potential benefits.
What’s more decentralized emergency control is
showing important advantages over centralized
control, especcialy with large data, calculation and
communication.
Several intelligent approaches have been
proposed for preventing large-scale emergencies. On
the one hand, there have been some previous
attempts to take advantage of agents and multiagent
systems as control systems (Lehnhoff 2011, Häger
2012, Panasetsky 2012, Negnevitsky 2008). On the
other hand, some different machine learning models
– including artificial neural networks (ANNs) – have
been successfully applied for power system security
assesment, as for example (Kalyani 2012, Voropai
2012, Niebur 1994).
The use of ANN models as a trigger system of
the multi-agent control systems let take advantage of
some of the properties of ANNs (such as pattern
recognition) and agents (reactivity, proactivity and
sociability) making preventing large-scale
565
Negnevitsky M., Tomin N., Panasetsky D., Haeger U., Voropai N., Rehtanz C. and Kurbatsky V..
Neural Multi-agent-based Approach for Preventing Blackouts in Power Systems.
DOI: 10.5220/0004906505650570
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 565-570
ISBN: 978-989-758-015-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
emergencies is more effective and reliable (Tomin
2013, Negmevitsky 2013) . The paper proposes a
neural multi-agent-based system for preventing
blackouts in power systems. This system includes
some experience and developments obtained at the
University of Tasmania (Australia), the Melentiev
Energy Systems Institute (Russia) and the TU
Dortmund (Germany) in developing intelligent
systems for a disaster management in modern power
systems.
2 PROBLEM DESCRIPTION
Several studies identified voltage instability as one
of the major reasons of blackouts (PSDP 2007,
Tomin 2013, CAMS 2008). A typical blackout
scenario develops as follows: high system loading
(due to heavy transfers across the grid) is followed
by events that initiate protection system actions. As
a result, some lines are disconnected, the grid
becomes even more overloaded, and consumption of
reactive power is increased, causing a cascading
effect in which voltages drop even further. Practical
experience demonstrates that most blackouts begin
with a large disturbance (a disturbance, which may
or may not cause cascading failures), which leads to
a slow deterioration of the system conditions (PSDP
2007, Tomin 2013).
Failures of protection and emergency control
devices as well as human errors are the two biggest
causes of large-scale blackouts. Most blackouts
begin with a large disturbance, which leads to a slow
deterioration of the system conditions. The system
parameters may still remain within specified limits,
but many of these parameters are on the boundary of
stability. If such conditions are identified as pre-
emergency, preventive actions can be taken, and
major events avoided.
Unfortunately, in current competitive
environment, such conditions may not be easily
detected because different problems may
simultaneously occur in different parts of a large
network within different jurisdictions. The
liberalisation process in power systems has created
an additional interface which can adversely impact
communication and coordination activities between
operators on both sides.
Multi-agent models are oriented towards
interactions and collaborative phenomena. It is
perfect suitable for resolve so called the irony of
interconnected power grids that are owned by
separate and often competing companies. In this
case a technical cooperation between interconnected
grids to a certain extent militates against the pure
profit motive.
3 PROPOSED SYSTEM
The proposed system consists of two main parts: the
alarm trigger, which is an intelligent neural network-
based system for detecting possible alarm states in a
power system, and the competitive–collaborative
multi-agent control system (MACS).
3.1 The Hierarchy Multi-agent Control
System
The innovation here in using a decentralized
structure in which distributed “agents” operate in
either competitive or collaborative modes,
depending on the system security state, so that fast
and robust responses can be provided in both normal
and emergency conditions – responses directly
tailored to the very different needs of each of these
two conditions. Agents are hardware or software
entities operating in virtual or real environments,
and will be distributed among all serial devices in a
power system – generators, transmission lines,
transformers, and power flow controllers (PFCs).
The MACS is a hierarchy of agents located at
different levels (Fig. 1):
1) The top-level agent (Advisor) – the objective is
to initiate a collaborative protocol of agents
located on the middle level, setting up their
priorities, and coordinating their actions. In
practice this might be a joint security center of
several TSOs.
2) Middle-level agents (the transmission system
operator-level) – the objective is to initiate
control actions according to the goals set by the
Advisor. These actions include PFCs between
different systems.
3) Low-level agents (i.e. device-level agents –
generators, transmission lines, transformers and
loads) – the objective is to achieve the goals set
by the respective middle-level agents within their
jurisdictions. The low-level agents are
specialized devices responsible for specific areas
of power generation, transmission and
distribution.
The Advisor receives messages from middle-level
agents about the current state of the interconnected
system, and if required proposes appropriate actions
to control power flows between different systems. If
the Advisor receives an alarm message from the
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
566
Figure 1: Block diagram of the MACS.
security alarm system, it assesses the severity of the
situation, and if required takes control over the
middle-level agents.
For normal conditions, we organise competitive
control by the middle-level agents as follows:
optimal power flow (OPF) is used to determine the
optimal settings of power-flow controlling devices
in the area of responsibility of each agent. Using
PFC devices as an example, representing them by
phase-shifting transformers and flexible alternating
current transmission system (FACTS) devices
(Häger 2012). The PFCs were installed to increase
transmission capacity and controllability of the grid.
In normal operation conditions, each transmission
system operator (TSO) used OPF methods to
optimise settings of their PFC devices, to reduce
internal congestions as required by market rules. The
objective function is to minimise the generation
costs by optimising power flows according to the
market situation.
However, in emergency conditions, all TSOs will
need to coordinate their PFC devices to stabilise the
system. We achieve this coordination through the
use of MACS. In the collaborative mode, the
objectives of the agent operation changes: a middle-
level agent seeks and receives help from the low-
level agents that belong to the neighbouring middle-
level agents. For example, under an emergency in
System A due to voltage instability, a middle-level
agent A will redefine the objective function of low-
level agent B1 of System B (Systems A and B are be
connected via a tie transmission line) to increase
reactive power input from the neighbouring system.
3.2 a Neural Multi-agent-based
Approach
In order to distinguish between competitive and
collaborative mode, we need to overcome an issue of
identifying pre-emergency conditions. This paper is
concerned with the real time identification of alarm
states that are dangerous for the system security. We
examined a clustering approach based on the self-
organized Kohonen neural network. The Kohonen
alarm trigger identifies pre-emergency conditions
using enormous amounts of data with incomplete
and distorted alarm patterns and activates the MACS
(Fig. 2).
Multi-agent control system
Figure 2: Diagram of a neural multi-agent-based system.
The security alarm system is trained using a set of
training examples based on randomly generated
events in a power system. The clusters is identified
using test cases representing a set of normal and
emergency conditions in the power system. As a
result, a Kohonen ANN-based clustering system is
able to classify power system states in real time and,
if required, to produce an alarm. The main objective
is to rank power system states with respect to their
potential for causing voltage instability.
The Kohonen network is trained off-line and
used on-line to classify the system operating state
based on the patterns created in the off-line mode.
The Kohonen network is divided into power system
states as follows: normal, alarm, emergency
(correctable) and emergency (non-correctable).
Here, a normal state implies that all parameters of
the power system are maintained within specified
normal operation limits:






(1)







. 1,2,…,
(2)



|
|
|

|
,1,2,…,
(3)
NeuralMulti-agent-basedApproachforPreventingBlackoutsinPowerSystems
567




for every branch 
(4)
where

is the real power generation at bus ,
is
the total system demand,

is the total real power
loss in the transmission line,
|
|
is the voltage
magnitude at bus ,

represents the real power
flow at branch ,
and
being the number
of generators and the number of buses in the power
system, respectively.
Real-time measurements are used to assess the
system state. Kohonen network-based monitoring
provides a warning when the system security is
under threat (Fig. 2).
4 CASE STUDY
The proposed neural multi-agent-based system was
implemented in JADE (Java Agent Development
Framework). The Kohonen security clustering
model is realized in STATISTICA 6.0. MATLAB
and Power System Analysis Toolbox are used as
modeling tools. In this paper, we demonstrate the
proposed approach on the modified IEEE One Area
RTS-96 power system. Active power flow is
directed from Subsystem B to Subsystem A.
Subsystem A is a low-voltage distribution subsystem
being in stressful conditions because of reactive
power shortage, which potentially may cause voltage
instability. Subsystem B is a high-voltage
transmission subsystem with surplus of reactive
power.
The modified system has 53 buses and dynamic
elements to represent generators and loads. In
Subsystem B, there is an excess of reactive power
produced by reactors at busses 107, 111, 113.
Subsystem A has a deficit of reactive power. In
Subsystem A, the sources of reactive power are
Non-controlled Reactive Power Sources (NRPS) –
capacitor. Each load is modeled as exponential
recovery load. The exponential recovery load model
can adequately represent the load behavior during
voltage instability.
To demonstrate the proposed approach, the test
system is subjected to the following sequence of
disturbance: t=10 s – the loss of transformer T101-
208. We assumed that two types of automatic
control can be used in the power system:
Conventional Automatic Control System (CACS)
(includes TGs, AVRs and OXLs on each generator,
and OLTCs on transformers connected to buses
204–210), and MACS (Fig. 3) – in addition to the
set of local controllers, it includes OLTCs on
transformers connected to busses 101, 102 and 103.
Figure 3: Subsystem A with installed Load Agents and
Generator Agents (the device level agents).
4.1 Conventional Automatic Control
Modeling
The loss of transformer T101-208 immediately leads
to an overload of generators connected to bus 201.
After about t=15s, the OLTC starts to change the
transformation ratio for boosting the secondary
voltage at the load. This leads to a gradual overload
of all generators in the subsystem. At t=300 s, the
rotor current limits are exceeded on all generators as
the system does not have any reactive power
reserves. At t=500 s, the primary voltage at the load
reaches 0.8 p.u. due to the OLTC actions and
insufficient reactive power, while the secondary
voltage is maintained close to the nominal (Fig.4).
Figure 4: The system voltage profile under the CACS
control.
In this case, the AVR fails to secure critical voltage
levels in the primary network, and at about t=500 s,
cascading voltage decrease takes place. As a result,
after t=600 s, stator currents of the generators
increase rapidly (Fig. 5), and the voltage at generator
busses decrease even further. This leads to the
disconnection of generators by their protection
systems, and as a result, the system voltage collapses
(Fig. 4).
0 100 200 300 400 500 600 700 800
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Load Primary and Secondary Voltages, p.u.
U205
U401
U204
U402
U208
U403
U209
U404
U206
U405
U210
U406
U207
U407
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
568
Figure 5: Stator current and voltage profiles under the
CACS control.
4.2 Neural Multi-agent Automatic
Control Modeling
The MACS coordinates main controllers of reactive
power in the system in order to prevent voltage
instability. The MACS detects dangerous levels of
excitation currents of a number of generators and
blocks the OLTCs on transformers.
The Kohonen network-based security alarm
system uses the following inputs: voltages at busses
204 – 210 (primary voltages); voltages at busses 401
– 406 (secondary voltages); AVR excitation voltages
for generators G301 – G309; OXL output signals for
generators G301 – G309 and stator currents of
generators G301 – G309. Fig. 6 represents a
topological map of the Kohonen network. The
network is trained off-line to identify clusters
corresponding to the following operating states of
the power system: normal (cluster A); alarm 1–5
(cluster B); emergency 1 and 2 (correctable) (cluster
C); and emergency 3 (non-correctable) (cluster D).
Figure 6: The Kohonen topological map.
After the loss of transformer T101-208, generators
connected to bus 201 are overloaded (Figs7,8).
When the multi-agent scheme is available, as soon
as the Kohonen network detects the alarm state at
time t=10 s, the MACS is activated in order to
prevent the system from further deterioration.
Figure 7: The system voltage profile under the MACS
control.
Figure 9: The sequence of messages between agents in the elimination of emergency.
0 100 200 300 400 500 600 700 800
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
1.05
1.1
time,sec.
U,p.u.
Stator Voltages, p.u.
U
G
303 U
G
306 U
G
307
0 100 200 300 400 500 600 700 800
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
I,p.u.
Stator Currents, p.u.
I
G
303 I
G
306 I
G
307
0 50 100 150 200 250 300 350 400
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
1.02
1.04
Load Primary and Secondary Voltages, p.u.
U205
U401
U204
U402
U208
U403
U209
U404
U206
U405
U210
U406
U207
U407
NeuralMulti-agent-basedApproachforPreventingBlackoutsinPowerSystems
569
Figure 8: The system stator current and stator voltage
profiles under the MACS control.
The local automation reduces the AVR setting. GAs
at busses 202 and 203 begin to increase reactive
power output until the generator excitation currents
reach their near-critical values. Collaborative actions
of GAs and LAs are allowed to unload G301-303
(Fig. 9).
As a result, the system voltage profile improves,
as can be seen in Fig. 7, and the Kohonen network
does not detect any deterioration at t=130.63 s. From
t=206 s, the Kohonen network identifies the normal
state, however, the alarm 3 state is also still activated
because the system is still in the normalization of
post-emergency state.
Thus, as a result of the MACS control actions,
the subsystem can maintain its stability without load
shedding via coordinating available sources of
reactive power.
5 CONCLUSIONS
This paper proposes a neural multi-agent-based
approach to the system monitoring and control with
the goal of identifying potential voltage instability
problems before they lead to major blackouts. The
proposed MACS structure is hierarchical; it consists
of the top-level agent, middle-level agents and low-
level agents. Under normal operating conditions, the
MACS operates in a competitive mode; low-level
agents exchange information with other agents to
maintain their local conditions within specified
limits and to maximize profits of their respective
companies. An alarm state triggers a collaborative
mode in which the agents coordinate their actions to
prevent a system blackout.
REFERENCES
IEEE PES PSDP Task Force on Blackout experience,
mitigation, and role of new technologies, blackout
experiences and lessons, Best practices for system
dynamic performance, and the role of new
technologies, IEEE Special Publication 07TP190, July
2007.
Wang X., Shao W., Vittal V., 2005. Adaptive corrective
control strategies for preventing power system
blackouts, in Proc. of 15
th
PSCC, Liege.
Lehnhoff S., Häger U., Zimmermann T., Rehtanz C.,
2011. Autonomous distributed coordination of fast
power flow controllers in transmission networks, in
IEEE PES ISGT 2011 Europe, Manchester, UK.
Häger U., 2012. Agent-based real-time Coordination of
Power Flow Controllers, Dissertation, TU-Dortmund
University.
Panasetsky D.A., Voropai N.I., 2009. A multi-agent
approach to coordination of different emergency
control devices against voltage collapse. IEEE
Bucharest PowerTech.
Negnevitsky M., 2008. Computational Intelligence
Approach to Crisis Management in Power Systems,
Int. Journal of Aut. and Control, 2 (2/3), 247-273.
Kalyani S., Shanti Swarup K., 2012. Design of pattern
recognition system for static security assessment and
classification, Pattern Analysis & Applications, vol.
15, 299-311.
Voropai N., Kurbatsky V., Tomin N., Panasetsky D.
“Preventive and emergency control of intelligent
power systems,” In Pros. of PES ISGT Europe, Berlin.
Niebur D. 1994. Kohonen self-organizing neural network
for power system security assessment, Thesis #1244,
Lausanne, EPFL.
Negnevitsky M., Tomin N., Panasetsky D., Kurbatsky V.,
2013. Intelligent Approach for Preventing Large-Scale
Emergencies in Electric Power Systems, IEEE
PowerTech 2013, Grenoble, France.
Negnevitsky M., Voropai N., Kurbatsky V., Tomin N.,
Panasetsky D., 2013. Development of an Intelligent
System for Preventing Large-Scale Emergencies in
Power Systems, IEEE/PES GM, Vancouver, BC,
Canada.
IEEE PES CAMS Task Force on Understanding,
Prediction, Mitigation and Restoration of Cascading
Failures “Initial Review of Methods for Cascading
Failure Analysis in Elect. Power Trans. Syst.,” In
Proc. IEEE PES GM, Pittsburgh, PA USA July 2008.
0 50 100 150 200 250 300 350 400
1
1.02
1.04
1.06
1.08
1.1
1.12
1.14
t,сек.
U,отн.ед.
Напряжение Статора, отн.ед.
U
G
303 U
G
306 U
G
307
0 50 100 150 200 250 300 350 400
0.7
0.8
0.9
1
1.1
1.2
I,отн.ед.
Ток Статора, отн.ед.
I
G
303 I
G
306 I
G
307
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
570