Page 9 - 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
5. Regression task
a time delay between the end of the input window and the prediction - so if input is the window [t−0.2 ms,t], we predict the R∆φ signal at time T = t + 0.2 ms. The network also performed well on this regression task, achieving MSE error < 10−5 on the validation data. The ability to predict the future excluded flux radius demonstrates that this sys- tem, or a similar machine learning model, could be used to anticipate the occurrence of staircase instability in real- time. Sample predictions of the R∆φ signal may be seen in figures 8 and 9.
The foregoing work served to establish a network structure that performs well on the classification task (i.e. agrees with the output of an edge detector). We used this network struc- ture as a starting point for a network that can perform regres- sion on the actual value of R∆φ over time. We created this regression task dataset in the same way as the dataset for the classification task, except that the variable to be predicted is the radius at midplane value. In addition, we introduce
Figure 9.
Two additional shots, plotted as in figure 8.
9