IEEE Power & Energy Magazine - March/April 2018 - 82

as shown in Figure 6. Upon the successful completion of the
training, there is no need to apply power flow analysis to
every system state to select the contingency. in other words,
dnn can provide a justification of the system status based
on the input data. the proposed whole process is illustrated
in Figure 7.

System Fault Diagnosis
although both focus on maintaining secure operation,
system fault diagnosis differs from security assessment
in that the former deals with an outage that has already
occurred while the latter makes predictions of the future
contingency. during real-time power system operation,
the voltage/current values change with load fluctuations,
which implies different operating conditions of lines and
buses. therefore voltage/current values can be used as an
intuitional signal of a fault. to train a dnn for fault diagnosis, historical data can be applied as the input, and the
output includes the fault type and the associated possibil-

PG
QG
PD
QD
V
θ
Input: Measurement/
Simulation Data

DNN Training

figure 6. Training the DNN.

ity. in this case, the dnn works as a classification tool to
pin the fault location. With the assistance of a well-trained
dnn, a system fault can be quickly spotted based on the
observation of a current or voltage abnormality, even in the
face of a large-scale transmission system with hundreds or
thousands of branches. this intelligent fault diagnosis process is depicted in Figure 8.

Cascading Outage Prediction
the cascading outage prediction differs from security assessment in that it is a multiperiod risk-assessment procedure,
in which an initial disturbance would trigger a sequence of
dependent component outages in future time intervals. this
in turn adds an even heavier computational burden to the system operator. to crack this computational intractability, if we
assume that the procedure of cascading outages is markovian,
then all the possible outages can be formed as a tree structure
and traversed using an mctS, as shown in Figure 9.
in the markovian tree structure, each node stands for a
system state, and each branch represents an outage. Similar to the
previous two studies, historical
data and simulation runs can be
Secure
used to perform the dnn training
State
for the policy network. then, for
Alarm
any system state, a risk index can
State
be assigned to evaluate its severity, which is calculated as the sum
Insecure
State
of the immediate cost after the
initial outage and the expectation
Output: System
of the risk indices of the subtree
Status
outages. this can be used for evaluation in the value network.

Supervised Learning
(Based on Historical Data)
Establish the Policy Network:
Identify the Potential
Contingencies via Measured
Data Under Different Scenarios

Reinforcement
Learning
(Based on Simulation)

Scenario Tree
Generation:
Construct the Scenario
Tree to Represent the
System Operation States

Establish the
Value Network:
Get the Security
Index of Each
System State Via
Power Flow
Calculation
Input: System
State Variables

Model Test:
Input Real-Time Data into
DNN to Conduct Online Security
Assessment and to Further
Implement Corrective Control
and Stabilization Measures

Secure
Alarm
Insecure
Ranking

Output:
Severity
Evaluation

Deep Neural Network Training:
Use the Sampled System States to Decide
the Number of Hidden Layers, the Neurons in
Each Layer, and the Weights for Each Neuron
Connection for an Accurate System
State Classification

figure 7. A DNN-based online security assessment.
82

ieee power & energy magazine

march/april 2018

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - March/April 2018