Neural-Network Behavioural Modelling and Digital Predistortion of RF Power Amplifiers

Power amplifiers in 4G / 5G base stations are non-linear near saturation: they spread energy into the neighbouring channels (poor ACLR / ACPR) and distort the constellation (EVM). From real MEASURED PA input / output baseband I/Q (OpenDPD datasets of TU Delft: an auxiliary PA with 5 x 20 MHz LTE and a Doherty PA with 10 x 20 MHz 64-QAM) the project builds behavioural models of the PA - the classical memory polynomial and generalised memory polynomial (least squares) and GRU / LSTM / TCN networks (PyTorch on the CPU) - and digital predistorters: MP / GMP by indirect learning and TCN / GRU by direct learning through the frozen neural PA model. The TCN is the best PA model on both PAs (test NMSE -36.68 dB APA, -34.42 dB DPA, versus -31.11 / -31.58 dB for the best memory polynomial). Through the learned PA model the predistorters lower the ACLR from -27.80 to -48.68 dBc (APA, GRU-DLA) and from -31.05 to -47.90 dBc (DPA, MP-ILA x3); the Doherty 64-QAM EVM falls from 4.05 % to 1.00 % (TCN-DLA). These DPD results are simulations through a learned PA model, not hardware measurements. GNU Octave, ngspice and OpenDPD itself were run; a complete MATLAB version (written, not run) and a web application with five databases are included.

Category: Electronics and Communication Engineering (RF and Microwave Engineering, Power Amplifier Linearisation, Digital Predistortion, Memory Polynomial, Generalised Memory Polynomial, Indirect and Direct Learning, GRU / LSTM / TCN Neural Networks, ACLR / ACPR / EVM, GNU Octave, ngspice, OpenDPD, MATLAB, Python)

Electronics & Communication Engineering Project E09 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.