Page 8 - Detection and prediction of a beam-driven mode in field-reversed configuration plasma with recurrent neural networks

Nucl. Fusion 60 (2020) 126025 C. Scott et al
Figure 8. Regression accuracy for predicting excluded flux radius at midplane for two shots of C-2U, with a time delay between network input and predicted values. Actual in red, prediction in blue. The regression LSTM model is able to use the signals described in section 2 to predict staircase with a lead time of 0.2 ms. As discussed in section 6, this is a large enough prediction window to include this machine learning model in a control loop to mitigate staircase instability.
while others do not converge. A preliminary result is that the magnetic probes (both radial and axial) seem to have high predictive power, whereas the bolometry array consistently ranked low-accuracy This latter result is somewhat surpris- ing, as bolometry is one of the main diagnostics used for human diagnosis of the staircase instability: During the micro- burst both the magnetic and bolometer signals show a strong, coherent n = 2 spatial mode, but the perturbation amplitude is much larger on the bolometers, making them better a pos- teriori diagnostics for the human eye. See figure 5 for an
example of both signals at the moment of plasma instability. In this figure we can clearly see that the instability is easy to pick out on the bolometry array after the fact, but the mag- netic probes demonstrate it more clearly prior to the onset of instability. Bolometry data is therefore a poor predictor of staircase, whereas the magnetic probe signal may be used to predict (and therefore steer away from) staircase instability. This may explain why the neural network relies more on mag- netic probe information than on bolometry. We intend to ana- lyze these trends more precisely in future work.
8