Digital Predistortion Model Development for Broadband Radio Frequency Power Amplifiers

The Power Amplifier (PA) is an essential part of modern wireless transmitters, but it is also a source of nonlinearities and memory effects that degrade signal quality and reduce spectral efficiency. Digital Predistortion (DPD) corrects these artefacts. This thesis presents two behavioural modelling frameworks for DPD designed to lower computational complexity and simplify hardware implementation. The first framework utilises conventional polynomial-based models, including the Generalised Memory Polynomial (GMP), Decomposed Vector Rotation (DVR), Magnitude Selective Affine function (MSA), and Instantaneous sample indexed MSA (IMSA). In these models, their main computational load arises from multiplicative cross-terms that combine the magnitudes of current and delayed input samples. The proposed method quantises input magnitudes into power-of-two levels and replaces costly multiplications with constant-time shift operations. This substitution reduces reliance on Digital Signal Processor (DSP) units, Look-Up Table (LUT), and pipelining resources in Field Programmable Gate Array (FPGA) hardware. Quantisation granularity is configurable to balance model accuracy and hardware complexity. The approach maintains the models’ linear-in-parameter structure and enables coefficient extraction by Least Squares (LS) regression. Thus, the quantised models maintain similar modelling fidelity while lowering latency and power consumption. The second framework uses a hybrid architecture that combines the temporal modelling of a complex-valued Echo State Network (ESN) with the GMP basis functions. A large, sparse reservoir with fixed random recurrent weights maps the baseband signals into a rich dynamical feature space. The reservoir is initialised with a spectral radius below 1 to ensure the Echo State Property (ESP), which makes the reservoir states dependent only on the input sequence, not the initial conditions. This property is verified through temporal recurrence patterns, bifurcation analysis, and consistency tests. To handle the reservoir’s high dimensionality, complex Principal Component Analysis (PCA) extracts the principal components. The transformed states are concatenated with the GMP feature matrix and the input samples. The output coefficients are extracted by ridge-regularised LS regression, which performs convex optimisation and avoids expensive Backpropagation Through Time (BPTT) algorithm. This design allows fast training and stable inference with limited hardware resources. Tests on Orthogonal Frequency Division Multiplexing (OFDM) datasets with baseband bandwidths of 60 MHz, 100 MHz, and 200 MHz confirm the effectiveness of both frameworks. The bit-shifted cross-term models provide a low-cost inference path with only slight increases in error. It makes them practical for hardware with limited resources. Moreover, the GMP-ESN hybrid model offers an alternative way to reduce computational load while reliably capturing nonlinearities and memory effects. It effectively tracks the PA characteristics and suppresses spectral regrowth across the bandwidths.