Automatic Piecewise Extreme Learning Machine-Based Model for S-Parameters of RF Power Amplifier

Abstract

This paper presents an automatic piecewise (Auto-PW) extreme learning machine (ELM) method for S-parameters modeling radio-frequency (RF) power amplifiers (PAs). A strategy based on splitting regions at the changing points of concave-convex characteristics is proposed, where each region adopts a piecewise ELM model. The verification is carried out with S-parameters measured on a 2.2–6.5 GHz complementary metal oxide semiconductor (CMOS) PA. Compared to the long-short term memory (LSTM), support vector regression (SVR), and conventional ELM modeling methods, the proposed method performs excellently. For example, the modeling speed is two orders of magnitude faster than SVR and LSTM, and the modeling accuracy is more than one order of magnitude higher than ELM.