Model predictive control of plasma profilesat DIII-D

neural network has been developed to predict future values of plasma profiles in real time using past data and a set of proposed actuator inputs. Thispredictive model offers many opportunities for real time control. The simplest method, which is already possible with the current model, is to simply make multiple predictions using different proposed actuator inputs, and selecting the inputs that lead to the ``best'' predicted profile. This can be improved upon by modifying the model to make over a longer horizon, sinstead of just predicting the next timestep, it can predict the next timesteps. This then allows for longer lookaheads in the control action, which allows us to predict and preemptively mitigate instabilities before they become dangerous, to potentiallyfind new routes to H-mode, which currently requires very strong heating, and can be difficult to achieve.