A multivariate analysis of disruption precursors on JET and AUG G.Sias 1, R. Aledda 1, B. Cannas 1, R. S. Delogu 2, A. Fanni 1, A. Murari 2, A. Pau 1, the ASDEX Upgrade Team 3 and JET Contributors 4 1 Electrical and Electronic Engineering Dept. - University of Cagliari, Cagliari, Italy 2 Consorzio RFX (CNR, ENEA, INFN, University of Padova, Acciaierie Venete SpA), Corso Stati Uniti 4, 35127 Padova, Italy 3 Max-Planck-Institüt für Plasmaphysik - EURATOM Association, Garching, Germany 4 EUROfusion Consortium, JET, Culham Science Centre, Abingdon, OX14 3DB, UK
Comparative multivariate analysis The aim of the work is to perform a disruption predictor both for JET and AUG by combining the prediction capability of a set of precursors in common between the two machines. The prediction capability of the signals reported in the following has been investigated. Acronym q95 LM Pradtot Pfrac Vloop ne_fr li dwmhd/dt Zcc signal Safety factor at 95% of poloidal flux Locked mode signal Total radiated power Total radiated power/ Total input Power Loop Voltage Line-averaged plasma density / Greenwald limit Internal inductance Time derivative of the total energy Vertical position of plasma centroid G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 2
Databases JET-ILW, 116 flat-top disruptions occurred between 2011-2012 Data set shots shot range training 77 81867-83340 test 39 83341-83698 AUG W-wall, 102 flat-top disruptions occurred between 2011-2014 Data set shots shot-range training 68 26903-28812 test 34 28813-30130 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 3
AUG pdf of signals Acronym q95 LM Pradtot Pfrac Vloop ne_fr li dwmhd/dt Zcc Disruption predictor N N N G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 4
JET pdf of signals Acronym q95 LM Pradtot Pfrac Vloop ne_fr li dwmhd/dt Zcc Disruption predictor y N N G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 5
Performance indexes Premature detection (PD): the fraction of disruptions triggered too much in advance Successful prediction (SP): the fraction of disruptions correctly predicted Tardy detection (TD): the fraction of disruptions triggered too late Missed alarm (MA): the fraction of disruptions predicted as safe. AUG PD SP TD t D -0.5 t D -0.002 t D MA Time [s] JET PD SP TD t D -1.5 t D -0.01 t D MA Time [s] G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 6
AUG signals as single predictors Training-set Signal Threshold PD SP TD MA LM [V] 0.15 0.13 0.74 0.01 0.12 Pradtot [MW] 12.0 0.12 0.75 0.09 0.04 Pfrac 1.60 0.16 0.84 0.00 0.00 Vloop [V] 2.61 0.06 0.66 0.06 0.22 li 1.52 0.00 0.79 0.04 0.16 Zcc [m] 0.085 0.10 0.16 0.03 0.71 Test-set Signal Threshold PD SP TD MA LM [V] 0.15 0.26 0.65 0.00 0.09 Pradtot [MW] 12.0 0.00 0.53 0.15 0.32 Pfrac 1.60 0.00 0.85 0.03 0.12 Vloop [V] 2.61 0.12 0.62 0.03 0.24 li 1.52 0.09 0.74 0.00 0.18 Zcc [m] 0.085 0.00 0.03 0.03 0.94 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 7
JET signals as single predictors Training-set Signal Threshold PD SP TD MA q95 3.80 0.19 0.36 0.01 0.43 LM [mt] 0.34 0.01 0.78 0.14 0.06 Pradtot [MW] 8.50 0.51 0.17 0.01 0.31 Pfrac 1.80 0.17 0.64 0.03 0.17 Vloop [V] 3.50 0.27 0.19 0.01 0.52 li 1.24 0.10 0.43 0.01 0.45 Zcc [m] 0.28 0.29 0.32 0.01 0.38 Test-set Signal Threshold PD SP TD MA q95 3.80 0.15 0.33 0.00 0.51 LM [mt] 0.34 0.03 0.69 0.13 0.15 Pradtot [MW] 8.50 0.54 0.23 0.00 0.23 Pfrac 1.80 0.23 0.59 0.05 0.13 Vloop [V] 3.50 0.23 0.21 0.03 0.54 li 1.24 0.08 0.41 0.00 0.51 Zcc [m] 0.28 0.18 0.23 0.08 0.51 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 8
Data fusion method for disruption prediction (*) For each signal a set of thresholds, which allows to achieve predefined PD rates, is identified A score (SC) is assigned to each threshold depending on the corresponding PD rate. At each time step during a discharge, the following steps are executed: 1. Each signal is compared with respect to its own thresholds and the corresponding score is assigned when a threshold is exceeded 2. The scores from the individual signals are totaled to form the aggregate score 3. A disruption warning is triggered when the aggregate score exceeds an optimized threshold value. The threshold values for each signal, their corresponding scores and the threshold on the aggregate score are optimized on the training set. *S.P. Gerhardt et al 2013 Nucl. Fusion 53 063021 doi:10.1088/0029-5515/53/6/063021 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 9
Data fusion method for disruption prediction AUG - Thresholds Signal PD=0.1 PD=0.06 PD=0.03 SC=2 SC=3 SC=5 LM [V] 0.16 0.2 0.27 Pradtot [MW] 13 14.5 17 Pfrac 2 2.3 2.6 Vloop [V] 2.59 2.61 2.9 li 1.45 1.465 1.48 Zcc [m] 0.085 0.0865 0.088 JET - Thresholds Signal PD=0.1 PD=0.06 PD=0.03 SC=1 SC=2 SC=4 LM [mt] 0.3100 0.3183 0.3300 Pradtot [MW] 22.66 22.96 23.95 Pfrac 2 2.2 4.7 Vloop [V] 5.7 6.1 6.97 Li 1.24 1.25 1.262 Zcc [m] 0.326 0.327 0.38 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 10
Data fusion method for disruption prediction TRAINING SETS: only variables in common between the two machines AUG JET PD SP TD MA ΔSP=SP-0.94 0.01 0.94 0.03 0.01 0.00 0.01 0.94 0.03 0.01 0.00 0.01 0.93 0.04 0.01-0.01 0.01 0.94 0.03 0.01 0.00 0.00 0.90 0.04 0.06-0.04 0.00 0.90 0.04 0.06-0.04 0.01 0.94 0.01 0.03 0.00 0.01 0.93 0.03 0.03-0.01 PD SP TD MA ΔSP=SP-0.79 0.06 0.79 0.04 0.10 0.00 0.04 0.82 0.04 0.10 0.03 0.06 0.78 0.04 0.12-0.01 0.06 0.79 0.04 0.10 0.00 0.04 0.77 0.04 0.16-0.02 0.06 0.79 0.04 0.10 0.00 0.06 0.48 0.01 0.44-0.31 0.04 0.82 0.04 0.10 0.03 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 11
Data fusion method for disruption prediction TEST SETS AUG PD SP TD MA 0.03 0.91 0.00 0.06 0.03 0.91 0.00 0.06 0.00 0.85 0.03 0.12 0.03 0.91 0.00 0.06 JET PD SP TD MA 0.08 0.74 0.03 0.15 0.05 0.77 0.03 0.15 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 12
Data fusion method for disruption prediction TEST set: 144 safe shots False alarm (FA): the fraction of safe shots predicted as disruptions Successful prediction (SP): the fraction of safe shot correctly predicted AUG FA SP 0.08 0.92 0.05 0.95 0.07 0.93 0.04 0.96 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 13
Locked Mode indicator Features extraction procedures in time and frequency domains: 1. Detrending and off-set removal of the normalized LM amplitude 2. σ of FFT on a 51.2ms sliding window N/2 i 3. f i abs FFT i on a 51.2ms sliding window 1) 2) 3) 4) G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 14
Locked Mode indicator Prediction performance: AUG Training set Signal Threshold PD SP TD MA LM_indicator 1.9E-04 0.06 0.84 0.04 0.06 LM 0.15 0.13 0.74 0.01 0.12 Test set PD SP TD MA 0.03 0.74 0.12 0.12 0.26 0.65 0.00 0.09 JET Training set Signal Threshold PD SP TD MA LM_indicator 1.1E-05 0.01 0.88 0.06 0.04 LM 0.34 0.01 0.78 0.14 0.06 Test set PD SP TD MA 0.05 0.77 0.10 0.08 0.03 0.69 0.13 0.15 G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 15
Comparison with the actual LM trigger AUG: the LM trigger on the considered data set achieves more than 0.50 of MA both on training and test sets JET: the LM trigger on the considered data set achieves SP=0.61 and SP=0.41 on training and test sets respectively. G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 16
Future work LM indicator will replace the LM signal as predictor input Others physical-based indicators will be added in order to improve the predictor performance For both machines, the database will be updated with most recent pulses (safe and disrupted) A performance statistical analysis taking into account disruption classes will be performed G. Sias et al. 1st IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis Nice 1 st June Page 17