Comparative Analysis of Resource-Constrained MP-LS and Neural Network DPD in Deep Saturation
22nd International Conference on Synthesis, Modeling, Analysis and Simulation Methods, and Applications to Circuits Design, SMACD 2026, Dresden, Almanya, 29 Haziran - 02 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/smacd70206.2026.11647791
- Basıldığı Şehir: Dresden
- Basıldığı Ülke: Almanya
- Anahtar Kelimeler: 5G OFDM, Digital predistortion, GaN HEMT, neural networks, power amplifiers, resource-constrained
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
The high peak-To-Average power ratio (PAPR) inherent in 5G and 6G OFDM-based modulation schemes forces power amplifiers (PAs) to operate close to their deep saturation limits. While this maximizes power efficiency, it inevitably triggers severe nonlinear distortion and hard clipping. Conventional digital predistortion (DPD) techniques, such as Memory Polynomials (MP), handle mild nonlinearities reasonably well. However, they struggle significantly under deep saturation unless equipped with high-order polynomial terms a feature that consumes more digital signal processing (DSP) resources on fieldprogrammable gate arrays (FPGAs). To tackle this hardware bottleneck, this paper introduces a resource-constrained DPD architecture based on an Augmented Real-Valued Time-Delay Neural Network (ARVTDNN). We deliberately constrained the feature extraction matrix, restricting the memory depth to M=2 and the polynomial order to K=3, thereby significantly reducing hardware complexity. We experimentally evaluated this neural network approach against a classical Least Squares-based MP (MP-LS) model, driving a GaN HEMT PA into deep saturation using a 100 MHz 5G OFDM signal. The results are positive. Thanks to its nonlinear activation functions, the neural network effectively compensates for the truncated feature space. Under the exact same stringent hardware constraints, the neural network DPD achieved a Normalized Mean Square Error (NMSE) of-44.04 dB and an Error Vector Magnitude (EVM) of 0.63%, outperforming the classical MP-LS model's-40.09 dB and 0.99%. Ultimately, this demonstrates a highly robust, resource-efficient DPD solution suited for next generation wireless transmitters.