Estimation of Plasma Vertical Position by Long-Short Term Memory Network with Time2Vec in a Small Tokamak Device PHiX

Abstract

Time evolution of plasma vertical position is estimated by using long-short term memory networks (LSTM) with Time2Vec technique which incorporates temporal information into a neural network. Since many tokamak devices have elongated cross-section in achieving high performance whereas accurate vertical position feedback control is required in order to avoid vertical displacement events (VDEs). Our data-driven model, using experimental data obtained from a small tokamak device PHiX in Tokyo Institute of Technology, can estimate the plasma vertical displacement by incorporating operational scenario coils current data. The model achieved high performance by combining Time2Vec with LSTM. We can also interpret the weights extracted from a trained, data-driven model by comparing the model’s predictions.