AI-powered Cross-Layer Optimization for Weather-Adaptive UDWDM-Based Hybrid FSO-FTTX Communication Systems

Abstract

Ultra-dense wavelength division multiplexing (UDWDM) free-space optical (FSO)-fiber-to-the-x (FTTx) hybrid networks encounter significant performance challenges under dynamic atmospheric conditions, which limits their reliability for next-generation applications in 5G/6G networks. This study presents a random forest-based machine-learning framework for real-time cross-layer prediction of optical signal-to-noise ratio (OSNR), quality factor (QF), and bit error rate (BER) in weather-impaired scenarios. A synthetic dataset comprising 4,800 instances was generated via numerical simulations of UDWDM-FSO/FTTx networks using four modulation schemes: On-Off Keying (OOK-NRZ), Quadrature Phase Shift Keying (QPSK), Polarization Division Multiplexed QPSK (PDM-QPSK), and Polarization Shift Keying (PolSK), input power levels ranging from -20 to 5 dBm, FSO link lengths up to 12.5 km, and diverse weather conditions, including clear skies, fog, haze, and rain. The Random Forest (RF) model achieved exceptional predictive accuracy across various metrics, including OSNR (R² = 0.995345; MAE = 0.133931; RMSE = 0.186337), QF (R² = 0.999955; MAE = 0.010762; RMSE = 0.015132), and BER (R² = 0.999085; RMSE = 1.65e-6; MAE =5.37e-07). Shapley Additive Explanations (SHAP) feature importance estimation identified weather and FSO link length as the most dominant variables affecting network performance. Statistical testing confirmed the existence of significant differences between predicted and ground-truth values (p < 0.05), demonstrating the model's robustness in diverse weather conditions. This framework supports adaptive power allocation and modulation switching strategies to enhance network resilience against adverse weather effects in intelligent optical communication systems optimized for future wireless technologies, such as fog- or rain-resilient backhaul and fronthaul in smart cities and IoT ecosystems.