Wideband RF power amplifier predistortion using real-valued time-delay neural networks

This paper suggests the application of real-valued time-delay neural networks (RVTDNN) for power amplifier (PA) behavioral modeling and linearization. The weights of the RVTDNN model and linearizer are identified using the backpropagation learning algorithm (BPLA), which is applied to the measured input and output signals of the PA. The RVTDNN scheme is first successfully used to accurately predict the dynamic nonlinear behavior of a 250 W LDMOS Doherty amplifier driven with 4 carrier (4C) WCDMA signal. The RVTDNN is then applied for the construction of a digital predistortion to improve the linearity of the same Doherty amplifier. The ACPR of the linearized Doherty amplifier revealed an adjacent channel power ratio (ACPR) of better then 50 dBc at the three offset frequency (5 MHz, 10 MHz, 15 MHz).