LED Model-Driven Reservoir Computing for Signal Equalization in OFDM-VLC System

This paper introduces a digitally realized, model-driven Reservoir Computing (RC) equalizer for Orthogonal Frequency Division Multiplexing (OFDM) in Visible Light Communication (VLC) systems. The design incorporates a Light Emitting Diode (LED) model building upon the frameworks of Time-Delay Reservoir Computing (TDRC) and Physics-Informed Neural Networks (PINN). In contrast to traditional RC, which uses a randomly connected neural network reservoir, the proposed method integrates a physical model that emulates the nonlinear electro-optical behavior of LEDs into the reservoir layer. This approach directly incorporates the intrinsic characteristics of LEDs, enabling the equalizer to more effectively capture the signal distortions caused by VLC channels. As a result, the model-driven RC significantly improves signal recovery under varying channel conditions, including different bandwidths and modulation voltages. Experimental results show that the proposed method achieves up to 56.0% and 34.6% reductions in Bit Error Rate (BER) compared to traditional RC and Volterra equalizers, respectively. The method also improves average subcarrier Signal-to-Noise Ratio (SNR) by 1.2 dB over Volterra and by 1.5 dB over traditional RC, resulting in an achievable rate improvement by nearly 10%. Additionally, the model-driven RC offers a key advantage in terms of data efficiency for the training stage, requiring only 20% of the total dataset compared to the 50% required by traditional RC and Volterra equalizers.