Particle swarm optimization of an online trained repetitive neurocontroller for the sine-wave inverter

The paper presents an evolutionary optimization of a novel neurocontroller for the single phase sine-wave inverter. The controller is trained in online mode. The adaptation algorithm takes into account repetitiveness of the process to be controlled. The cost function evaluates performance of the controller over the whole period of the reference signal and the weights are updated only once per period of this signal. A model-free concept is implemented and hence no prior identification of the plant is needed. The controller employs the backpropagation algorithm for updating its weights in order to adapt to changing load conditions. The controller is nonlinear and the tuning procedure involves determining a good set of values for at least three parameters: a number of neurons, an error gain and an output gain. In our recent publication devoted to the development of the control algorithm we relied on guessing and checking at the tuning stage. Here the particle swarm optimizer (PSO) is used to find the optimal set of values for these parameters. The gradientless search algorithm enables a designer to work with any user-defined performance index that reflects desired system behavior. The effectiveness of the proposed approach is illustrated with the help of numerical experiments. The controller tuned by the PSO is capable to maintain a high-quality output voltage waveform in the presence of the periodic disturbance caused by nonlinear loads.