Prediction of Discharge Current using Neural Network in Hall Thruster

Abstract

We have been developing a prediction code of discharge current using neural network for constructing auto-controlling system in Hall thrusters. The neural network is feedforward neural network, which consists of 5 layers with 100 neurons. We adopted backpropagation method to the network and updated weights by AdaGrad. We used training 25500 data sets that consists of operation condition (inner and outer coil current, xenon mass flow rate, discharge voltage and time) and discharge current. The code could predict unknown discharge current history within relative error 1% with three days. The relative error with 2250 training data sets remains less than 1% within eight hours calculation on a standard PC. Considering actual operation, it is necessary to make learning speed up.