Feasibility study of a neural network current controller for a boost rectifier.

Abstract

During the past two decades, Quality of Supply has become a serious problem for Variable Speed Drives in the industrial and commercial sectors. Quality of Supply problems can trip Variable Speed Drives, which results in loss of production, which is a significant problem in the paper and pulp industry. Researchers have proposed that Quality of Supply problems can be minimised in-house, using controlled front end rectifiers (boost rectifier), to maintain a regulated DC link voltage in the Variable Speed Drive configuration, as most faults are created by a varying supply voltage. This thesis extends the work performed on boost rectifiers by investigating the feasibility of replacing the classical controllers with a Continual Online Trained Artificial Neural Network current controller. The approach adopted in this thesis was to evaluate and extend the work previously performed on conventional boost rectifier current controllers and Continual Online Trained Artificial Neural Network current controlled inverter, at the University of Natal. During this evaluation, the respective controller shortcomings were identified and addressed. Thereafter the Continual Online Trained Artificial Neural Network current controller was modified, according to the control requirements of the boost rectifier, and used as a replacement for the conventional current controller in the boost rectifier system. Finally, the Continual Online Trained Artificial Neural Network current controller was evaluated to assess its viability as a current controller for a boost rectifier. The concept of implementing the real-time Continual Online Trained Artificial Neural Network current controller using a DSP (Digital Signal Processor) was described, along with the main features and practical limitations of existing commercial DSP's. It is shown that at the time of writing of this thesis, the commercially available DSP' s are not powerful enough to implement the Continual Online Trained Artificial Neural Network current controller. However this thesis also shows that it is feasible to implement the real-time controller on the newly released TMS320C67 DSP card.