Training recurrent neural network vector controller for inner current-loop control of doubly fed induction generator

Abstract

This paper proposes a novel recurrent neural network (RNN) based vector control method for a doubly fed induction generator (DFIG) and especially focuses on how to train the neural network controller for the current-loop control of DFIG. The proposed RNN vector control utilizes the stator voltage oriented frame and the role of the RNN is to substitute the two decoupled current-loop PI controllers in the conventional vector control technique. The objective of RNN training is to approximate optimal control and the RNN controller was trained by Levenberg-Marquardt (LM) algorithm. Forward Accumulation Through Time algorithm for the DFIG was developed to calculate Jacobian matrix needed by LM algorithm. Performance evaluation shows that the well-trained RNN controller has a very strong ability of tracking references under situations such as quickly rapid change reference and rotor parameter change.