Accurate rectifier models under modulated RF input signal are essential for Simultaneous Wireless Information and Power Transfer (SWIPT) design, yet conventional linear or polynomial models often fail to capture rectifier memory effects. Although previous research LSTM-based model improves accuracy, its training cost grows rapidly when many average input power levels must be supported. This paper proposes a scalable architecture composed of a shared cascade-LSTM network common to all power levels and power-level-specific fully connected networks. The shared LSTM is trained using input–output voltage waveforms from only a subset of power levels. For each target level, intermediate sequences are generated by the shared network and then used to train the corresponding fully connected subnetwork. Experiments with 64-QAM signals confirm accuracy comparable to the prior all-level training approach, while significantly reducing total construction time; the time savings increase as the number of supported power levels grows.
抄録全体を表示