Clustering and Fitting to Reduce PAPR in Multi-User OFDM Systems

In this paper, we provide a machine learning (ML) assisted method for reducing peak-to-average power ratio (PAPR) in realistic frequency-division multiplexing (FDM) scenarios of an orthogonal frequency division multiplexing (OFDM) system. The multi-user mode dictates a non-constant error vector magnitude (EVM) value in the frequency domain because of the variation in allocated bandwidth, modulation order, and power ratio for each user. This issue complicates the PAPR reduction using conventional approaches like clipping and filtering (C&F). Using tone reservation (TR) techniques, PAPR compensation signal can be distributed following the EVM value. However, the TR requires unreasonably lengthy hyperparameter computation. We suggest efficient clustering and ML-based hyperparameter calculation algorithms to address this issue. We demonstrate the trained clustering and prediction not only outperform C&F but also efficiently reduce PAPR to the lowest value defined by a TR technique with parameters discovered through brute-force search.