HyperLSTM: Hypernetwork Based Digital Predistortion for Wide Band Power Amplifier Linearization

This paper proposes HyperLSTM, a novel hypernetwork-based digital predistortion (DPD) architecture for power amplifier linearization. Unlike conventional LSTM DPD with fixed weights, HyperLSTM employs a secondary network that generates context-dependent modulation vectors to adapt the main LSTM’s weights dynamically. Experimental validation using 200 MHz 64-QAM OFDM signals on a Doherty power amplifier demonstrates that HyperLSTM achieves -52.00 dB adjacent channel power ratio (ACPR), outperforming baseline LSTM DPD by 10.0 dB and showing 21.77 dB improvement over the raw PA output. The proposed architecture achieves -45.5 dB normalized mean square error (NMSE) and 0.55% error vector magnitude (EVM) while adding only 18 additional parameters. These results represent 8.7 dB NMSE improvement over LSTM DPD and 25.0 dB NMSE improvement over no DPD. HyperLSTM offers superior nonlinear compensation for wideband communication systems, highlighting its potential for efficient and adaptive DPD in non-stationary operating conditions.