On-line Trained Neural Speed Controller with Variable Weight Update Period for Direct-Torque-Controlled AC Drive

The paper investigates further improvements of an adaptive ANN (Artificial Neural Network)-based speed controller employed in a DTC-SVM (Direct Torque Controlled - Space Vector Modulated) drive. An on-line trained ANN serves as a speed controller and does not need a process model to predict future performance. In comparison to the previously published solution, auto-adjusting ability has been added to the controller. The recurrent feedback inside the neural controller has been also introduced. Adaptive behaviour manifests in robustness to moment of inertia variation greater than 10 times. This feature is achieved by the learning algorithm running during system operation. Mentioned variable update period refers to one of the parameters connected with learning algorithm, namely frequency of calling backpropagation procedure (weights update procedure). Proposed control algorithm has been tested in simulation and verified experimentally. The behaviour of the drive has been compared to the one with previously proposed ANN-based speed controller with fixed settings of training algorithm.