On the use of LSTM-based estimation components for tokamak gas actuator control

• Drifts in the performance of the gas introduction modules used on JET were noticed. • The piezo actuator valve was modelled using a Long Short-Term Memory model. • Modelling was done in a Markov fashion. • Performance comparison was done with two linear Kalman filters. During the operation of JET, an approach using a Long Short-Term Memory (LSTM) neural network was explored to provide a data-driven means of developing and deploying a control system component. The LSTM was trained to model the piezo actuator valve which has non-linear hysteresis effects and required occasional recalibration due to drifts. A further motivation was to develop a process to make use of the large amount of control system data to enable control system performance improvements whilst minimising the use of machine time. The LSTM is developed in the form of a first-order Markov model where it takes inputs from pressure sensors around a gas reservoir and a control signal to the gas valve and predicts the next state of the downstream pressure sensor. A comparison of the LSTM time series model is done with two different linear Kalman filters.