In this paper, we consider the optimization of the performance of QPSK and 16-QAM coded orthogonal frequency division multiplexing (COFDM) signals over the non-linear and mobile satellite channel. A high power amplifier and Rician flat fading channel produces non-linear and linear distortions; an adaptive predistortion technique combined with turbo codes will reduce both types of distortion. The predistorter is based on a feedforward neural network, with the coefficients being derived using an extended Kalman filter (EKF). The conventional turbo code is used to mitigate Rician flat fading distortion and Gaussian noise. The performance over a non-linear satellite channel indicates that QPSK COFDM followed by a predistorter provides a gain of about 1.7 dB at a BER of 3 × 10-3 when compared to QPSK COFDM without the predistortion scheme and 16-QAM COFDM provides a gain of 0.5 dB output back-off and 1.2 dB signal to noise ratio at a BER of 3 × 10-5 when compared with an adaptive predistorter based on the Harmmerstein model. We also investigate the influence of the guard time interval and Doppler frequency effect on the BER performance. When the guard interval increases from 0 to 0.125T samples and the normalized Doppler frequency is 0.001, there is a gain of 0.7 and 1 dB signal to noise ratio at a BER of 6 × 10-4 for QPSK and 16-QAM COFDM, respectively.
- Recurrent neural networks