Home Search Collections Journals About Contact us My IOPscience

Disruption forecasting at JET using neural networks

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2004 Nucl. Fusion 44 68

(http://iopscience.iop.org/0029-5515/44/1/008)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 132.239.1.231
This content was downloaded on 14/05/2017 at 08:57

Please note that terms and conditions apply.

You may also be interested in:

A prediction tool for real-time application in the disruption protection system

B. Cannas, A. Fanni, P. Sonato et al.

An adaptive real-time disruption predictor for ASDEX Upgrade

B. Cannas, A. Fanni, G. Pautasso et al.

Prototype of an adaptive disruption predictor for JET

A. Murari, G. Vagliasindi, P. Arena et al.

Automatic disruption classification at JET

B. Cannas, F. Cau, A. Fanni et al.

On-line prediction and mitigation of disruptions

G. Pautasso, C. Tichmann, S. Egorov et al.

Prediction of density limit disruptions on the J-TEXT tokamak

S Y Wang, Z Y Chen, D W Huang et al.

Unbiased and non-supervised learning methods for disruption prediction at JET

A. Murari, J. Vega, G.A. Rattá et al.

Forecasting disruptions in the ADITYA tokamak using neural networks

A. Sengupta and P. Ranjan

An advanced disruption predictor for JET tested in a simulated real-time environment

G.A. Rattá, J. Vega, A. Murari et al.
INSTITUTE OF PHYSICS PUBLISHING and INTERNATIONAL ATOMIC ENERGY AGENCY NUCLEAR FUSION
Nucl. Fusion 44 (2004) 68–76 PII: S0029-5515(04)71108-4

Disruption forecasting at JET using

neural networks
B. Cannas, A. Fanni, E. Marongiu and P. Sonato1
Dipartimento di Ingegneria Elettrica ed Elettronica, Universita’ di Cagliari, Cagliari, Italy
1
Consorzio RFX, Associazione Euratom-ENEA sulla Fusione, Padova, Italy

Received 26 July 2002, accepted for publication 23 October 2003

Published 5 December 2003
Online at stacks.iop.org/NF/44/68 (DOI: 10.1088/0029-5515/44/1/008)
Abstract
Neural networks are trained to evaluate the risk of plasma disruptions in a tokamak experiment using several
diagnostic signals as inputs. A saliency analysis confirms the goodness of the chosen inputs, all of which contribute
to the network performance. Tests that were carried out refer to data collected from succesfully terminated and
disruption terminated pulses performed during two years of JET tokamak experiments. Results show the possibility
of developing a neural network predictor that intervenes well in advance in order to avoid plasma disruption or
mitigate its effects.

PACS numbers: 28.52.Av, 52.55.Fa

1. Introduction disruptions have been classified on the basis of the main

event that drove the disruption, e.g. density limit (planned/not
Plasma disruption is a serious event in which the plasma planned, with/without additional heating systems like neutral
magnetic confinement in a tokamak is suddenly lost. beam injection), killer pellet, low, li , low, q, etc.
Disruption prediction is an important area of study due to the Therefore, in principle, it should be possible to construct
possible damage a disruption could cause in a large tokamak a disruption prediction system by detecting disruption
device. Also, following a disruption, long reconditioning precursors with suitable online diagnostics.
periods of the machine may be necessary in order to restore The disruption predictor can be used to avoid the
the optimal machine conditions. disruptions or mitigate the unavoidable ones as suggested
Disruptions have different physical causes most of which in [4].
have been identified but whose dynamics is not known in detail. In this paper, we use an artificial neural network (ANN),
Therefore, present experimental operation frequently relies on fed by signals from a large number of plasma diagnostics, to
using human experience, and often the disruption cannot be estimate the risk of disruption. The data for network training
avoided or predicted and mitigated at all. and validation were selected from the JET database.
In some cases [1, 2], disruption dynamics and causes In particular, this study focuses on flat-top disruptions.
have been studied and explained from a physical point of During the flat-top phase the plasma is controlled in order
view. Disruptive events have been classified by analysing the to obtain a quasi-stationary plasma current, and steady
diagnostic signals recorded during the experiments that lead to equilibrium position and shape. Such sustained flat-top
scenarios are of importance because they will be the normal
a disruption. The dynamics of a disruption can be subdivided
operating conditions in next step tokamaks such as the
into four phases [2]:
International Thermonuclear Experimental Reactor (ITER). To
• the initiating event, which can be externally driven (e.g. a meet its objectives of producing significant fusion power ITER
control equipment failure), or internally driven (e.g. due will have much higher stored magnetic and thermal energies,
to high radiated power); than JET. These extrapolations in size and physics performance
• the precursor phase, which depends on the specific provide major constraints on the design of ITER.
initiating event and ends with a pre-quench state; This paper is organized as follows: section 2 is a survey of
• the thermal quench; work on neural networks and their applications to disruption
• the current quench. prediction, while section 3 describes the details of the database
that has been built to generate the prediction system. In
In [3] all the disruptions analysed were characterized by an section 4 the neural model of the disruption predictor is
initiating event for the disruption, with subsequent precursors, discussed. Finally, sections 5 and 6 summarize the results
which were recognizable in the diagnostic signals. Moreover, and conclusions.

0029-5515/04/010068+09$30.00 © 2004 IAEA, Vienna Printed in the UK 68

 Disruption forecasting at JET using ANNs

2. Disruption prediction with ANNs 2.2. Network learning

An ANN is a system composed of simple processing elements We used the Matlab Neural Network Toolbox to train the MLP
operating in parallel. The processing ability of the network network.
is stored in the inter-unit connection strengths (weights), The Levemberg–Marquardt training algorithm [6] gave
obtained by a process of adaptation to a set of training patterns the best performances in terms of error. It implements
(learning). the Levenberg–Marquardt variation of Newton’s method to
Today, ANNs are applied to an increasing number of real- modify the ANNs weights. The algorithm approximates the
world problems of considerable complexity. They offer ideal Hessian matrix as
solutions to a variety of classification problems such as pattern
H(w) ∼
= 2 · JT (w) · J(w) (5)
recognition, speech, character and signal recognition, as well
as functional prediction and system modelling. ANNs may and the gradient can be computed as
also be applied to control problems, where the input variables
are measurements used to drive an output actuator, and the g = JT · e, (6)
network learns the control function.
where J(w) is the Jacobian matrix containing first derivatives
2.1. Neural architectures of the network mean square error with respect to weights and
e is the vector of network errors. The Newton-like weights
There are many different types of ANNs. In this paper, a update can be written as
traditional multi layer perceptron (MLP) has been trained to
estimate the disruption risk from diagnostic data. wt+1 = wt − (JT (wt ) · J(wt ) + µt · I)−1 · gt . (7)
An MLP network consists of n0 inputs, one or more hidden
layers of neurons and one output layer with nL outputs. The In this way the training procedure approaches second-
neurons in each layer are connected with all the neurons of the order training speed without having to compute the Hessian
previous layer. The output of the ith neuron in the lth layer, matrix. Indeed, the Jacobian matrix can be computed through
xil , is a nonlinear function of the weighted sum of the previous a standard backpropagation that is much less complex than
layer outputs; computing the Hessian matrix. When µt = 0 equation (7)
  implements the Newton method with the approximate Hessian

nl−1
matrix, while a large value of µt implements the gradient
xi = f 
l
wij xj  ,
l l−1
(1) descent method.
j =1 To ensure good out-of-sample generalization perfor-
where i = 1, . . . , nl ; l = 1, . . . , L; nl is the number of neurons mances we used early stopping [7]. In this technique, the
in the lth layer, wijl is the connection weight between neuron available data is divided into three subsets:
j in the (l − 1)th layer and neuron i in the lth layer, f is a • The first subset is the training set, which is used for
nonlinear, usually sigmoidal, function and L is the number of computing the gradient and updating the network weights
layers. and biases.
The connection weights are determined by applying the • The second subset is the validation set. The error on the
network to a set of actual input–output values (the training set) validation set is monitored during the training process.
and comparing, through the error function, the network output The validation error will normally decrease during the
xL and the desired output d. The error function is the mean initial phase of training, as the training set error does.
square error E defined by However, when the network begins to overfit the data,
N nL (p) (p) 2 the error on the validation set will typically begin to rise.
p=1 j =1 (xj − dj ) When the validation error increases for a specified number
E= , (2)
nL · N of iterations, the training is stopped, and the weights and
(p) biases at the minimum of the validation error are returned.
where dj is the actual value of the j th output in the pth
• The test set is not used during the training, but is used to
example, nL is the number of outputs and N is the number of
compare different models.
examples.
Backpropagation learning updates network weights in the
direction in which the mean square error E decreases most 2.3. Related studies
rapidly, i.e. the negative of the gradient. One iteration of the Disruptions are a significant operational issue for tokamaks.
algorithm can be written as The literature reports various methods and approaches to
wt+1 = wt − αt · gt , (3) disruption prediction using neural networks: all of them are
based on the identification of either the initiating events or the
where gt is the gradient at the current iteration, and αt is precursors that lead to the disruption. As described in [1],
the learning rate. Higher performance can be obtained using a disruption can be schematized as a sequence of instabilities
second derivatives also, approximating Newton’s rule that could eventually end with an m = 2 MHD instability. The
wt+1 = wt − H−1 possible evolutions of a disruption sequence are different both
t · gt , (4)
in the underlying physics and in their dynamics. For example,
where Ht is the Hessian matrix of the mean square error at the the low q disruptions are sudden, there is no radiation induced
current values of weights [6]. contraction and there are no precursor oscillations or minor

69
B. Cannas et al

disruptions. On the other hand, density limit disruptions are ahead. The MLP showed good results in predicting all the
more complex than the low-q , involving a sequence of many input signals with K = 400.
events [1]. The main drawback of this approach is in the limited
Methods and approaches presented below constitute the number of disruptions that can be predicted using the selected
state of the art of the disruption prediction problem: a measurements. Indeed, the diagnostic signals used contain
description of the developed techniques is useful in depicting only a few types of precursors.
the advantages and limits of each approach.
One of the approaches that are present in literature relies 2.3.2. Disruption proximity predictors. This type of
on the physical limitations of a tokamak and on the theoretical predictors differs from those previously described by the output
stability limits of the maintenance of the plasma: the yielded: instead of using actual measurements as the target to
Greenwald plasma density limit, the Troyon high-β limit, the be predicted, an artificial output is assigned to directly detect
li –q diagram and the ratio of the radiated power to the input the event under investigation.
power, are some of the well-known physics-based techniques. In [3], an ANN detector for the MARFE formation
However, each of the listed techniques has to be fine-tuned has been developed. Indeed, most of the disruptions
for the specific machine configuration: e.g. the li –q diagram are characterized by the cooling of the plasma edge and
has been originally plotted for the limiter configuration; accompanied by a MARFE in the X-point region. The ANN
therefore, it has to be adapted to single null plasmas [1]; the output is defined as the probability of MARFE growth, which
Troyon high-β limit does not accurately describe the disruption would assess a forthcoming risk of disruption. Results were
boundary for individual discharges [8]. These limitations have improved in [4], where the ANN-predictor yields as an artificial
been overcome by alternative techniques; many of them are output target, the time to disruption, i.e. it estimates the time
ANN-based. interval up to the disruption. A threshold-based alarm is
Most of the ANN techniques are based either on applied on the basis of the time needed to take a countermeasure
predicting diagnostic signals, i.e. the model forecasts actual that will avoid or, at least, mitigate the disruption effects.
measurements that are known to be an indicator of incoming The system was implemented and tested for real-time
disruptions, or on predicting the proximity of the plasma to mitigation, and the obtained results show better detection
disruption by using artificial outputs. performances than those in [3].
However, in [4], the authors highlight the deterioration
of network performances on online tests, due to the slight
2.3.1. Diagnostic signal predictors. In the first case, one or
difference between the real-time signals and the stored ones.
more plasma measurements are used as targets to be forecast by
Moreover, it has been shown that new experiments that belong
the ANN. The ANN produces a set of future values of plasma
to an operational region different from those used for training
diagnostic signals that can be used in conjunction to a physical
are not well classified by the on-line predictor, thus presenting
model to decide if a disruption is incoming.
the so-called ‘ageing’ of the ANN.
In [8] the Troyon expression for the high-β limit is
Further improvements have been assessed in [12] using a
refined by an ANN that estimates, on the basis of magnetic
Fuzzy approach to classify data sets selected from ASDEX-U
measurements of the DIII-D tokamak, the value of β that
into four nonoverlapping subsets. Four ANNs have been
identifies the boundary between stable and unstable plasma
trained with the four subsets and a radial basis function network
configurations.
decides which of them has to be used. The output yielded by the
In [9] two buffered MLPs have been developed to forecast system includes the time to disruption as well, but an assertion
Mirnov measurements that identify m = 2 MHD instabilities procedure has been used to limit false alarms (FA). The results
in the TEXT tokamak. The first buffered MLP is trained presented refer to a small test set; therefore, the disruption
to produce the Mirnov coil measurement at the time t using prediction system has to be validated on a wider portion of the
the past measurements at time steps t − τ, t − 2τ, . . . , ASDEX-U database.
t − Nτ , where τ is the sampling time of the diagnostic In addition, work presented in [13] has been performed
system. Subsequently, the ANN works in autonomous mode on flat-top JET scenarios characterized by a single null
(i.e. the output at time t is fed back for the calculation of plasma. This work presents MLPs trained to recognize
the output at time t + τ ) to obtain the prediction at time plasma disrupting patterns in advance. The softmax activation
t + Kτ . The second buffered MLP is trained to produce the function has been used for the output layer; thus, the output
Mirnov coil measurement directly at time t +Kτ using the past of the network is the posterior probability that the input
measurements t − τ, t − 2τ, . . . , t − Nτ . patterns are disrupted or successful. The selected inputs have
A similar approach has been continued by using soft x-ray been widely used for the detection of both theoretical and
measurements to predict the disruption three time steps earlier experimental limits of a tokamak, such as the high-β limit, the
than the best network fed with Mirnov coil measurements li –q stability diagram and the power fraction limit. Therefore,
alone [10]. the ANN acts as a boundary detection system between the
On the basis of these works, an application on the safe and unsafe operational spaces. The disruption database
ADITYA tokamak and improvement of the ANN architecture used to train the network consisted of 360 disruptions with
is presented in [11]. The authors used as delayed input vectors plasma current above 1 MA. They all belong to Mark II divertor
a combination of diagnostic signals from four poloidally configurations. The training patterns were built with two
distributed Mirnov probes, one soft x-ray and one Hα monitor. patterns for each disruption: the first taken at the disruption
The ANN predicts the same input variables, K time steps time and the second when the mode lock reached the limit

70
 Disruption forecasting at JET using ANNs

set during operation. For the last 116 pulses of the Mark II It is worth noting that stored data are undersampled from
campaign the MLP detected all 22 disruptions that occurred, the real-time signals in order to save storage space, and that
but only 11 of them, 100 ms before the triggering of the signals belonging to different systems have different sampling
pulse termination. During the same period of operations, times and periods.
the system would have stopped five of the 73 successful The database consists of selected temporal sequences for
pulses. These results showed an improvement over traditional each pulse. Each sequence consists of 20 consecutive samples,
physics-based techniques, but a non-negligible sensitivity with a sampling time of 20 ms. The sampling time has been
to FAs. chosen in order to allow the synchronization among different
systems: the online neural predictor will work by gathering
3. Relevant diagnostic signals and database selection signals from various sources; therefore, it has to wait for the
slowest system before the computation.
In the results presented here, only unprocessed data, which Let us define the disruption event time, tD , as the time
are available in real time, are used (e.g. data relying on instant at which the plasma current suddenly begins to quench.
equilibrium reconstruction is not used). On the basis of the The numerical identification of tD is done by selecting the first
previous experiences on disruption prediction [13, 1, 4], nine maximum of I˙pla (the time derivative of the plasma current). In
diagnostic signals have been selected, as shown in table 1. order to avoid the selection of samples belonging to the thermal
These diagnostic signals have proved to be suitable disruption quench phase, which is not useful for precursor identification,
predictors and classifiers. In fact, we omitted the last 40 ms before tD . Therefore, the 20 selected
samples for a disrupted pulse belong to the time window
• they refer to global measurements of the plasma
[tD − 440 ms; tD − 40 ms]. For a successful pulse, the samples
parameters; moreover, they were used in traditional
belong to a randomly chosen time window of 400 ms. Since
(model based) disruption avoidance systems;
the nine signals belong to different acquisition systems, it is
• they can be normalized in order to obtain machine
necessary to synchronize their values by interpolating them
independent parameters [4]; indeed, they are calculated
and resampling at the chosen sampling time. Interpolation has
on every tokamak machine and can be relatively easily
been done by using the Matlab function interp1.
normalized.
Nevertheless, as will be shown by the saliency analysis (see
4. The neural model
section 5), some of the signals are less sensitive to the dynamics
of the system.
To model the disruption risk, a neural network is used, where
Pulses for training and validation sets were selected
each input node is associated with a diagnostic signal. So the
between 1167 successful pulses and 701 disrupted pulses in
neural model has as input the nine values of the chosen signals
the pulse interval 42 760–48 382, produced at JET between
at time instant t. The single output is a real number between 0
September 1997 and April 1999. The operational space
and 1 representing the risk of disruption.
exploited in this period is well represented by the data
In fact, the network target for each disrupted pulse is a
set, which includes experiments with high Ipla flat-tops,
sigmoid in the time window of 400 ms so as to represent a
high additional heating power values, etc, all belonging to
greater risk near the disruption:
Mark II and Mark II-Gas Box divertor configurations. Thus,
the training space is wide enough to assure generalization 1
capabilities to the system. Note that, as suggested in [4], y(t) = , (8)
1 + exp(−(t − (tD − tα ))/τ )
the performance of the system deteriorates if the network
is applied to an operational region outside of the training where tD is the disruption time instant, tα is set equal to 280 ms
space. The criteria used to select the discharges for network and τ is the sampling time (20 ms). Equation (8) is plotted in
database were: Ipla > 1.5 MA, X-point configuration and flat- figure 1. The choice of tα has been made by a trial and error
top plasma current profile, resulting in 183 disrupted pulses procedure.
and 778 safe pulses. Plasma current values below 1.5 MA The network target for the safe pulse is zero at every time
were discarded because these generally have little impact on instant.
subsequent conditioning and operation of the device. The training set consists of 86 disrupted pulses and 400
successful pulses, while the validation set consists of 35
Table 1. List of diagnostic signals used as input to the ANN (A.U.: disrupted pulses and 246 successful pulses. The test set
adimensional units). consists of 61 disrupted pulses and 132 successful pulses. Input
Signal name Units data have been scaled in the interval [0, 1] by dividing each
input by its maximum absolute value.
Plasma current Ipla [A]
In order to avoid local minima entrapments, multiple
Mode lock [T A−1 ]
Total radiated power Prad [W] training sessions have been carried out, with random initial
Plasma density ne [m−3 ] weights for each training in the session.
Total input power Pinp [W] For each session, a performance measure (as described
Plasma internal inductance li [A.U.] later) is computed in order to evaluate the best network of the
Stored diamagnetic energy derivative Ẇdia [W] session. Then, a new session is started with further tuning
Safety factor at 95% of minor radius q95 [A.U.]
Poloidal beta βp [A.U.] of the network parameters (i.e. the number of neurons in the
hidden layer, and the number of hidden layers). The training

71
B. Cannas et al

1
Network Target

0.9

0.8

0.7

0.6
Network Target

Disruption Event
0.5

0.4

0.3

0.2

0.1

0
0 50 100 150 200 250 300 350 400
Time (ms)

Figure 1. Network output for a disrupted sequence.

phase is iterated until a satisfactory performance has been Table 2. Performance results for the best MLP: FA = false alarm,
achieved. MA = missed alarm, TS = training set, VS = validation set.
Network performances have been quantified in terms of TS VS
false and missed alarms (MAs).
MA 3/86 1/35
For any pulse the ANN gives as output a time series of 20 FA 0/400 1/246
values representing the disruption risk. Detection of successful
and disrupted pulses is carried out as follows:
Table 3. Performance results on the test set for the best MLP.
• For a successfully terminated pulse, the diagnostic system
Test set
triggers a FA if the maximum value of the series is greater
than a predetermined threshold. Conversely, the pulse is MA 10/62
correctly classified. FA 0/132
• For a disrupted pulse the output series is analysed in the
time window [tD − 440 ms, tD − tP ms], where tP is Table 4. Performance results for the best MLP: FA = false alarm,
the prediction interval before the disruption instant tD . If MA = missed alarm.
the maximum value of this series is less than the threshold tP (ms)
value, the diagnostic system misses the alarm (MA). 40 60 80 100 120 140 160 180 200
Conversely, if this value is greater than the threshold,
Training set
the diagnostic system correctly triggers the alarm at least
MA (86) 2 3 4 3 4 5 8 9 17
tP ms before the disruption.
FA (400) 0 0 0 0 0 0 0 0 0
The threshold th is chosen by minimizing the detection error Validation set
expressed as MA (35) 0 0 1 1 2 3 4 5 5
FA (246) 0 0 0 1 1 1 1 1 2
ED (th, tP ) = PFA(th, tP ) · wFA + PMA(th, tP ), (9) Test set
where PFA(th, tP ) is the percentage of FAs, PMA(th, tP ) MA (62) 9 11 13 10 13 17 18 21 22
is the percentage of MAs, and wFA is a FAs weight factor. FA (132) 0 0 0 0 0 0 0 0 0
PFA is defined as the number of successfully terminated
pulses detected as disruption divided by the number of
successfully terminated pulses. that, in order to minimize the MAs, as will be crucial in future
PMA is defined as the number of disrupted pulses detected fusion reactors, it will be sufficient to use a different network
as successfully terminated pulses divided by the number of threshold modifying wFA .
disrupted pulses.
For an experimental machine, such as the JET tokamak, 5. Results
stopping the machine during a successful plasma discharge,
must be avoided [13, 14]. Thus, in order to consider FAs more In this work, we set tP = 100 ms, since this time interval
seriously than MAs, the misclassification of good pulses has should be sufficient to undertake, in advance, some adequate
been penalized by a weight factor wFA = 4. It is worth noting mitigation strategy.

72
 Disruption forecasting at JET using ANNs

1.2
Network Output
Plasma current (A)

1 6
2.5×10

0.8

Plasma Current (A)

2.0×106
Network Output

0.6 Desired detection time

Threshold
0.4

6
1.5×10
0.2

0
1.0×106

-0.2
17.5 18 18.5 19 19.5
Time (s)

Figure 2. MA on disrupted pulse 46800: network raises the alarm too late, for t = tD .

-10
5.0×10
Mode lock indicator (T/A)
Plasma current (A)
-10
4.5×10
2.5×106
-10
4.0×10

3.5×10 -10 Desired detection time

Mode Lock Indicator (T/A)

Threshold

3.0×10 -10 2.0×106 Plasma Current (A)

2.5×10 -10

-10
2.0×10
6
1.5×10
-10
1.5×10

1.0×10 -10

5.0×10 -11 1.0×106

0.0×10 0
17.5 18 18.5 19 19.5
Time (s)

Figure 3. MA on disrupted pulse 46800: even the mode locking signal does not raise the alarm.

The best network configuration is composed of nine have not been stopped by the ANN alarm. To evaluate the
inputs, two hidden layers with six and five hidden neurons, network prediction capability, the time interval tP has been
respectively, and one output, resulting in 89 network swept in the range [tD − 40 ms; tD − 200 ms]. Indeed, while
parameters. 100 ms seem to be sufficient to intervene in the experiment
Results obtained during training and validation are shown to mitigate the disruption effect, further delays could be
in table 2. As can be seen, the results are quite good, with introduced both by the diagnostic systems that produce the
2.8% of the MAs and 0.4% of the FAs in the validation set. input signals to the ANN, and by the neural computation
Table 3 shows the results for the test set with time itself. Table 4 summarizes the results which have been
tP = 100 ms. The successful pulses have not been stopped obtained with the same network, by re-computing PFA and
by the ANN alarm. Nevertheless, the performance deteriorates PMA and minimizing equation (9) with the new tP .
slightly: in 16% of the disruptions the MLP raises the alarm too Although the MA percentage is higher, the network
late (in the interval [tD −100 ms, tD ]), and the successful pulses does not trigger later than the mode lock indicator

73
B. Cannas et al

1.6 6
2.6×10

1.4 6
2.4×10
Network Output
Plasma current (A)
1.2
2.2×106

1
6

Plasma Current (A)

2.0×10
Network Output

0.8
1.8×106
0.6

Threshold 1.6×106
0.4

6
1.4×10
0.2

6
0 1.2×10

-0.2 1.0×106
18 18.5 19 19.5 20 20.5 21 21.5 22
Time (s)

Figure 4. FA on successful pulse 46314: the network alarm would have stopped the experiment.

1.6 5.0×106
. Output
Network
Wdia (J/s)
1.4

0.0×100
1.2

Stored Energy Derivative (J/s)

1
6
Network Output

-5.0×10
0.8

0.6
-10.0×106
Threshold
0.4

0.2 6
-15.0×10

6
-0.2 -20.0×10
18 18.5 19 19.5 20 20.5 21 21.5 22
Time (s)

Figure 5. FA on successful pulse 46314: ANN output is highly influenced by Ẇdia .

(Ml /Ipla > 3 × 10−10 T A−1 ), which is presently used in the In figure 5, the ANN output for the same pulse is plotted
on-line disruption protection system. together with the stored diamagnetic energy derivative, Ẇdia .
Network output and plasma current Ipla for a MA (pulse The behaviour of the diagnostic signal seems to justify the
46800) are shown in figure 2. The horizontal line indicates network response. When Ẇdia suddenly becomes large and
the network threshold, while the short vertical line indicates positive due to ELMs, the ANN output increases.
the desired detection time (tD − 100 ms). As can be noted, the However, the relation between the ANN output and Ẇdia
disruption was triggered only at t = tD . No precursor could be is not so straightforward, and a more complex relationship
identified by a visual inspection of the nine diagnostic signals including other inputs exists. Indeed, the ANN output does
used as input to the network. Also the mode lock signal did not always trigger corresponding with the Ẇdia decrease, as
not detect the immininent disruption, as shown in figure 3. can be noted for t close to 19 s.
In figure 4 a FA is presented for the successfully terminated Figure 6 shows the detection of a disruption for the
hot ion H-mode pulse 46314. disrupted pulse 43380. In this case, the network alarm occurred

74
 Disruption forecasting at JET using ANNs

1.2 6
2.0×10

1 Network Output
Plasma current (A)
1.8×106

0.8

Plasma Current (A)

6
Network Output

1.6×10
0.6

Threshold
0.4
6
1.4×10

Desired detection time

0.2

6
1.2×10
0

6
-0.2 1.0×10
17.8 18 18.2 18.4 18.6 18.8
Time (s)

Figure 6. Correct alarm on disrupted pulse 43380: network raises the alarm well in advance, and an early rise of the alarm suggests that
precursors are present more than 400 ms before the major disruption.

1.0×10 -9 1.2

1
8.0×10 -10 Mode Lock Indicator (T/A)
Network Output

0.8
-10
Mode Lock Indicator (T/A)

6.0×10
Desired detection time Network Output
0.6
-10 Threshold (both network and
4.0×10 Mode Lock Indicator)
0.4

-10
2.0×10
0.2

0.0×10 0
0

-2.0×10 -10 -0.2

17.8 18 18.2 18.4 18.6 18.8
Time (s)

Figure 7. Correct alarm on disrupted pulse 43380: network raises the alarm well before the mode locking trigger, which is set at 300 pT A−1 .

more than 400 ms before t = tD and well before the mode lock Results (table 5) show that the performance slightly
alarm (figure 7), which detects the disruption only 100 ms in deteriorates by setting zero to the stored diamagnetic energy
advance. derivative, while setting zero to the inputs like q95 or li yields
poor results on the trained network. Therefore, the presence of
5.1. Saliency analysis such signals is very important for the improvement of ANNs
for disruption prediction.
The saliency analysis is derived from the need to assess the
relative importance of different neural network inputs in terms 6. Conclusions
of the effect that each item has on the error function [5]. The
technique consists of leaving out one input variable after the An MLP has been successfully trained to forecast disruptive
other, and then recalculating the error function associated with events for JET, up to 100 ms in advance. Pulse samples have
the remaining variables. been selected in a temporal window of 400 ms; for disrupted

75
B. Cannas et al

Table 5. Saliency analysis on both training and validation sets for F. Milani, the EFDA-JET Task Force M and the UKAEA staff
the best MLP. for the assistance and help given in developing this work.
Input FA–TS MA–TS FA–VS MA–VS
Full set 0/400 3/86 1/246 1/35
References
Less Ipla 370 1 230 1
Less Ml 26 14 10 11 [1] Wesson J.A. et al 1989 Disruptions in JET Nucl. Fusion 29
Less Pin 190 4 107 3 641–66
Less li 202 19 115 3 [2] Schuller F.C. 1995 Disruptions in tokamaks Plasma Phys.
Less Ẇdia 3 4 5 4 Control. Fusion 37 A135–62
Less q95 397 1 245 1 [3] Pautasso G. et al 1998 Causes, precursors and mechanisms of
Less Prad 3 30 1 15 disruptions in Asdex Upgrade ICPP&25th EPS Conf. on
Less βp 192 1 121 0 Control Fusion and Plasma Physics (Praha, 29 June–3
Less ne 3 18 2 15 July) vol 22C (ECA) pp 520–3
[4] Pautasso G. et al 2002 On-line prediction and mitigation of
disruptions in ASDEX Upgrade Nucl. Fusion 42
100–08
pulses the window is constituted by the last 400 ms of the [5] Abrahart R.J. and Kneale P.E. 2001 Investigating the role of
discharge. The large data sets used assess the performance saliency analysis with a neural network rainfall-runoff
and reliability of the ANN predictor, and good generalization model Comput. Geosci. 27 921–8
[6] Hagan M.T. and Menhaj M. 1994 Training feedforward neural
capabilities have been obtained on unknown input patterns. networks with the Marquardt algorithm IEEE Trans. Neural
Moreover, satisfying results have been obtained when the ANN Networks 5 989–93
predictor triggered the alarm 200 ms before the disruption. [7] Smith M. 1993 Neural Networks for Statistical Modeling
Saliency analysis has been performed to verify the necessity (New York: Van Nostrand Reinhold) ISBN 0-442-01310-8
of certain signals for the correct prediction of disruptive [8] Wroblewski D., Jahns G.L. and Leuer J.A. 1997 Tokamak
disruption alarm based on a neural network model of the
sequences; of the nine input variables it is found that Ipla and high-beta limit Nucl. Fusion 37 725–41
q95 are crucial, whereas Ẇdia and ne are the least important. [9] Hernandez J.V., Vannucci A., Tajima T., Lin Z., Horton W. and
To summarize, ANNs are good, fast and reliable McCool S.C. 1996 Neural network prediction of some
disruption predictors suitable for real-time application since classes of tokamak disruptions Nucl. Fusion 36 1009–17
the multispace state parameters that identify the operational [10] Vannucci A., Oliveira K.A. and Tajima T. 1999 Forecast of
TEXT plasma disruptions using soft x-rays as input signal
window of an experimental device are sufficiently defined, in a neural network Nucl. Fusion 39 255–62
and since reliable diagnostic signals can be provided to the [11] Sengupta A. and Ranjan P. 2000 Forecasting disruptions in the
online ANN. ADITYA tokamak using neural networks Nucl. Fusion 40
1993–2008
[12] Morabito F.C., Versaci M., Pautasso G. and Tichmann C. and
Acknowledgments The ASDEX-Upgrade Team 2000 Fuzzy-neural approaches
to the prediction of disruptions in ASDEX-Upgrade Nucl.
This work has been carried out under the auspice of the Fusion 40 1715–23
RFX(ENEA)-Euratom Agreement and was performed partly [13] Milani F. 1998 Disruption prediction at JET PhD Dissertation
University of Aston in Birmingham
under the European Fusion Development Agreement. The [14] Walker M.L., Scoville J.T., Johnson R.D., Hyatt A.W. and
data used was obtained in the framework of the JET Joint Lee J. 1999 Automated fault detection for DIII-D Tokamak
Undertaking. The authors would like to thank T. Hender, experiments General Atomics Report GA-A23246

76