LightGBM-Based Data-Driven Modeling of Directly Modulated Laser and Optoelectronic Oscillator-Assisted Remote C-RAN Architectures

We propose a data-driven modeling approach based on light gradient boosting machine (LightGBM) for remote cloud radio access network (C-RAN) architectures assisted by a directly modulated laser (DML) and an optoelectronic oscillator (OEO). The architecture incorporates a dual-loop OEO to provide a low-phase-noise 20 GHz microwave source, achieving a phase noise reduction of 27 dBc/Hz at a 10 kHz offset compared to conventional sources. Concurrently, it leverages the interaction between DML chirp and fiber dispersion to achieve a gain of up to 19.0 dB in the intermediate frequency (IF) upper band of a 40 GHz millimeter-wave signal when the transmission distance over single mode fiber (SMF) is increased from 5 km to 35 km. To model this complex system, an end-to-end multi-output LightGBM model is developed, using only three measurable parameters as inputs: the frequency and amplitude of modulating IF signal, and the length of single mode fiber (SMF) link, to predict the millimeter-wave spectrum. The experimental results show that LightGBM model achieves high accuracy in predicting millimeter-wave spectral profiles, with a mean absolute error (MAE) below 1.70 dB for modulating IF signal frequencies in the 11–68 MHz upper sideband and below 0.81 dB for those in the 68–120 MHz upper sideband, outperforming both the conventional analytical model and other mainstream machine learning based models. Furthermore, LightGBM model's rapid inference capability enables real-time system performance evaluation at only 0.43 milliseconds per spectrum prediction, far surpassing the analytical model's 2,164 milliseconds.