Neural networks for real time determination of radiated power in JET

This article describes the use of neural networks (NNs) for the on-line computation of the radiated power in JET. The NNs have been trained using a database of about 120 discharges, for which the emitted power had been calculated via tomographic inversion of JET bolometric signals. In addition to the bolometric data, elongation and triangularity have been used as input to the NN, since these provide useful complementary information. Dedicated NNs have been designed for the determination of the total radiated power, the power from the bulk, and from the divertor region. All the NNs have been tested with a set of about 30 discharges with positive results. Moreover, the NNs can operate at full sampling speed and are therefore suited to follow edge localized modes and other rapid phenomena. The sensitivity of the NNs to failures in the input signals has also been tested, proving their robustness. Their possible use in feedback applications is finally briefly discussed. (C) 2002 American Institute of Physics.