Voltage source inverter output waveform compensation using adaptive intelligent control.

摘要

A single-layer neural network-based voltage compensation technique which generates minimum-distortion sinusoidal output voltages from a three-phase PWM inverter used for uninterruptible power supplies (UPS) is described. The proposed compensation technique is implemented in a microprocessor-based controller constructed in the stationary d-q frame where the controller sampling rate is twice the inverter switching frequency. The structure of a feed-forward artificial neural network connects network inputs and outputs through multiple linear or nonlinear neuron models, and processes these input/output data associations in a parallel distributed manner. Network inputs in the form of UPS load voltage commands and load current feedback are propagated forward in the network each controller sampling period generating the inverter output voltage commands, the network outputs, which are converted to three-phase inverter switching signals using the space vector PWM waveform generation process. Each controller sampling period, the network weights are modified by the controller learning process with the objective of minimizing the cost function {dollar}{lcub}1over 2{rcub}{lcub}cdot{rcub}lbrackepsilonsb{lcub}rm dv{rcub}sp2 + epsilonsb{lcub}rm qv{rcub}sp2rbrack{dollar} where {dollar}epsilonsb{lcub}rm dv{rcub}{dollar} and {dollar}epsilonsb{lcub}rm qv{rcub}{dollar} are the measured d- and q-axis UPS load voltage errors. Once the cost function is minimized, the neural network input/output mapping approximates the PWM inverter/output filter circuit inverse transfer characteristics.; Three neural network-based controller configurations were studied via computer simulation: (i) A controller with one hidden layer and hyperbolic tangent squashing functions (six hidden layer nodes used) and the conventional backpropagation learning algorithm utilized. (ii) A controller without hidden layers (single-layer network) and the conventional backpropagation learning algorithm utilized. (iii) A controller without hidden layers and a modified backpropagation learning algorithm utilized.; The third controller configuration was incorporated into the design of an experimental controller. Three-phase UPS system experiments and simulations using this control technique were run using balanced passive, unbalanced passive, and nonlinear loads at inverter switching frequencies consistent with the use of GTO technology. Using the same load models and system parameters, three-phase UPS system simulations were run utilizing average load voltage control and the repetitive controller technique described in (3). Results obtained from these simulations were served as a basis for evaluating the performance of the neural network-based controller.; The application of the neural network-based inverter controller concept is intended for high-power three-phase UPS systems where inverter switching frequencies are reduced and a broad range of UPS loads and output filter designs are encountered. (Abstract shortened by UMI.)