Abstract: This paper proposes a multi-band universal digital predistorter model based on one-hot coding and longshort term memory (LSTM) neural network, which can effectively linearize broadband RF power amplifiers working in multiple frequency bands. By introducing different one-hot codes in the training set to represent signals of different frequencies, the trained neural network can be applied to power amplifier linearization of different bands without changing the network structure and model parameters. To verify the effectiveness of the method, an experimental platform with two RF power amplifiers working at 2. 6 GHz and 4. 9 GHz respectively is established. The normalized mean square error (NMSE) for predistortion nonlinear modeling can reach -40 dB for the two frequency bands. Then, a 100 MHz bandwidth 5G NR signal is used to verify the linearization performance of the method for the two-band power amplifiers. Experiments show that the multiband universal digital predistorter can improve the adjacent channel leakage ratio (ACLR) of the two power amplifiers by 19. 42 dB and 17. 91 dB respectively at 100 MHz below the center frequency, and 15. 73 dB and 15. 17 dB respectively at 100 MHz above the center frequency, which verifies the effectiveness of the proposed nonlinear model.