Research On Behavioral Modeling Of RF Power Amplifiers Based On Genetic Neural Network

With the rapid development of modern wireless communication, RF power amplifiers, as the most expensive, maximum power consumption as well as nonlinear device, have been widely researched in the RF domain, and the RF PA behavioral modeling is one of the central issues; Meanwhile, in recent years, neural networks and its optimization algorithms have been constantly applied in communication systems, this open a new window for the RF PA behavioral modeling research. In this thesis, the main researches include:(1) A study of the reference model constructed by real value time delay neural network(RVTDNN);(2) A study of the RF PA behavioral modeling based on genetic algorithms neural network(GANN);(3) A study of RF PA behavioral modeling based on improved genetic algorithms neural network(IGANN);(4) A comparative analysis between the literature’s behavioral model with the neural networks’ related models acquired by this article.This thesis mainly makes theoretical research from a mathematical point of view to study the RF PA nonlinear distortion characteristics and behavioral modeling. Actually, RF PA behavioral modeling is a process of mathematical modeling on RF PA’s nonlinear characteristics. The main contribution of the thesis can summarize as:1. GANN applied in RF PA behavioral modelingCombined the RVTDNN with the modern optimization algorithms, not only makes the model complexity in structure than the RVTDNN model to be reduced by more than half, but extracted parameters also appears to be more efficient with LM algorithm. The simulation results show the fitting accuracy has increased nearly 3dB.2. Research on RF PA behavioral modeling based IGANNSince the slow convergence and dual phenomenons occur by crossover and mutation operations of basic GA, an IGANN model has proposed, which exhibit more favorable performance in global optimization than GANN model. Simulation results show the fitting accuracy has enhanced about 4dB than GANN model under the similar structure complexity.Finally, two different application RF PAs have been used to simulate and varify the obtained behavioral models. Results show that the obtained GANN and IGANN behavioral models output not only fit very well with the actual measured output, but more better than with the RVTDNN model. Besides, in the case of large output power PAs, the neural network’s models exhibit a better performance compared with the polynomial model proposed by the literature, which has an important application in large output PA behavioral modeling.