Computation of optimal switching patterns for voltage-controlled inverters using neural-network software

The feasibility of neural networks for optimal voltage control of three-phase power inverters was investigated. The case study involved elimination of the 5th, 7th, and 11th harmonics of the output voltage using four switching angles per quarter-cycle of the output frequency. Regular and sparse neural networks were experimented with. The regular network requires training, but the number of neurons is low. Training is not necessary for the sparse network which can be synthesized using a simple formula. The sparse network also offers higher accuracy than the regular network, at the expense of a significantly higher neuron count. Results of computer simulations have confirmed the viability of the proposed technique.