Digital predistortion of RF Doherty power amplifier based on GWO-MA-BiLSTM model

Abstract

In this paper, a behavioral model integrating the grey wolf optimizer (GWO) and a multi-head attention (MA) based bidirectional long short-term memory (BiLSTM) network (GWO-MA-BiLSTM) is proposed for digital predistortion (DPD) of doherty power amplifiers (DPA) in radio-over-fiber communication systems. Unlike other BiLSTM based models, GWO-MA-BiLSTM significantly reduces model training time by leveraging the grey wolf optimizer algorithm to identify the optimal combination of network architecture and hyperparameters. The multi-head attention mechanism enables the model to selectively pay attention to the input signals at different moments. This enables the model to dynamically adjust the attention at different times, so as to better capture important features and dependencies, and improve the performance and accuracy of the model. To verify the modeling ability of the GMO-MA-BiLSTM model for the nonlinearity and memory effect of PA, this paper conducts behavioral modeling of the dynamic AM/AM characteristics, dynamic AM/PM characteristics and PSD of the DPA circuit. The experimental results show that the model prediction data is in good agreement with the circuit simulation data.