Dynamic LSTM for 5G Signal Power Amplifier Behavioral Model

The non-linear characteristics of power amplifiers (PAs) used in the 5G communication systems are becoming increasingly complex. The existing behavioral models of PAs have not considered the dependency changes of nonlinearity on the input signal, leading to unsatisfactory performance. This letter proposes two new recursive neural network units to simulate the characteristics of non-linear behavior of PAs with varying input signal power, namely, dynamic long short-term memory network (DLSTM) and its enhanced version multiplicative dynamic long short-term memory network (MDLSTM). The proposed network achieves dynamic changes in the learnable parameters of LSTM units at each time step, as opposed to traditional parameter sharing, to enhance the model's dependency on input signals. Experimental results indicate that these modifications enable better learning and representation of the nonlinear characteristics of PAs under 5G signal waveforms. Additionally, compared to traditional LSTM unit, the computational complexity remains almost unchanged.