Deep Neural Network Based Predistorter with ReLU Activation for Doherty Power Amplifiers

In this paper, we propose a deep neural network based DPD (DNN-DPD) where ReLU activation is used instead of sigmoid activation to avoid the gradient vanishing problem. The ReLU function is implemented using a half-wave rectifier that acts as the activation function for the hidden layers in the DNN-DPD. Experimental results using a Doherty PA will show that the proposed DNN-DPD can achieve an additional 3 to 4 dB of linearity compensation compared to a conventional Volterra-DPD or shallow NN. Furthermore, to achieve comparable performance to sigmoid activation, the number of multipliers can be reduced by a factor of 100. This confirms that the DNN-DPD approach has a large potential to overcome the complicated memory effects of Doherty PAs with reduced DPD complexity.