Free space optical （FSO） communication systems have emerged as an important complement to conventional fiber-optic and radio-frequency （RF） communication technologies due to their inherent advantages， including ultra-high bandwidth， high data transmission rates， low latency， and operation in license-free spectrum bands. These characteristics make FSO systems particularly attractive for medium- and short-range high-speed wireless transmission scenarios such as satellite communications， unmanned aerial vehicle （UAV） relays， emergency communications， and last-mile access networks. Despite these advantages， the practical deployment and stable operation of FSO systems are significantly challenged by their strong sensitivity to atmospheric conditions. In realistic propagation environments， FSO link performance is highly susceptible to dynamic and complex environmental factors， including atmospheric turbulence， fog， haze， rain， and dust. These effects introduce severe nonlinear distortions and random fluctuations in received optical signals， leading to time-varying attenuation， signal fading， and degradation of link quality. Consequently， transmission stability and reliability are difficult to guarantee， especially under rapidly changing weather conditions. To mitigate performance degradation， numerous adaptive optimization strategies have been proposed， such as adaptive modulation， power control， beam adjustment， and hybrid FSO/RF link switching. However， the effectiveness of these strategies fundamentally relies on the availability of accurate and real-time link quality assessment or prediction.Traditional physical channel models often require simplified assumptions and struggle to capture the highly nonlinear and time-varying characteristics of real-world FSO channels. Although machine learning-based methods have recently been introduced for link quality prediction， existing approaches still face notable limitations， including insufficient nonlinear fitting capability， limited prediction accuracy， slow convergence speed， and vulnerability to local optima. These shortcomings restrict their applicability in complex dynamic environments. Therefore， developing a robust and accurate Q-factor prediction algorithm with enhanced fitting capability， high prediction accuracy， and stable training behavior is of critical importance. This study aims to address these challenges by proposing an improved data-driven prediction framework for FSO link quality under complex dynamic weather conditions.This study proposes a Q-factor prediction algorithm for FSO links employing on–off keying （OOK） modulation， referred to as the Newton-Raphson-Based Optimizer Backpropagation （NRBO-BP） algorithm. The proposed method integrates the Newton-Raphson-based optimizer （NRBO） with a Backpropagation Neural Network （BPNN） to establish an accurate nonlinear mapping between physical link parameters and the corresponding Q-factor， thereby improving both prediction accuracy and training stability. First， multidimensional input features are constructed based on the physical characteristics of the FSO system. These features include transmit power， link distance， beam divergence angle， atmospheric refractive index structure constant， and atmospheric visibility under different weather conditions. Such a feature design ensures comprehensive characterization of both system-level and environmental factors influencing link quality. The output label is defined as the Q-factor calculated using a theoretical OOK-based FSO link model， ensuring data consistency and reliability during training. To overcome the limitations of conventional gradient-based optimization， the NRBO is employed to optimize the neural network parameters. The Newton–Raphson search rule utilizes second-order gradient information to precisely determine the weight update direction， enabling faster and more accurate convergence. Furthermore， when the curvature of the optimization landscape approaches zero or when continuous stagnation occurs along a certain search direction， a trap-avoidance operator is activated. This operator introduces adaptive perturbations that allow the optimization process to escape flat regions and local optima. By suppressing ineffective update paths and enhancing productive exploration， NRBO effectively reduces redundant updates， parameter oscillations， and training instability. After NRBO-based initialization and optimization， the neural network undergoes end-to-end training through standard gradient backpropagation. The optimized initial parameters significantly accelerate error convergence and improve robustness against channel dynamics， ultimately enhancing the predictive capability of the model in complex environments.Simulation results demonstrate that the proposed NRBO-BP algorithm exhibits significant advantages in predicting the Q-factor of FSO links under complex dynamic environmental conditions. Performance evaluation conducted on the complete test dataset shows that NRBO-BP achieves superior results across all evaluation metrics， including the coefficient of determination （R²）， mean absolute error （MAE）， and root mean square error （RMSE）. Specifically， NRBO-BP achieves an R² value of 0.985， indicating an excellent fit between predicted and actual Q-factor values. Compared with convolutional neural networks （CNN）， conventional BPNN， and random forest （RF） models， the MAE is reduced by approximately 68.17%， 69.43%， and 24.89%， respectively. Similarly， RMSE is reduced by approximately 65.54%， 69.1%， and 53.35%， respectively. These results confirm the strong adaptability and robustness of the proposed algorithm under varying atmospheric conditions. Further analysis of optimization performance reveals that NRBO provides faster convergence and more stable training compared to genetic algorithm （GA） and particle swarm optimization （PSO）. The final convergence fitness of NRBO is approximately 29.86% lower than that of GA and 14.13% lower than that of PSO， demonstrating its superior optimization efficiency. When combined with BP neural networks， NRBO-BP achieves approximately 3% improvement in R² compared to GA-BP and PSO-BP， while MAE and RMSE are consistently reduced. These results highlight the effectiveness of NRBO in enhancing convergence behavior and strengthening the predictive capability of neural networks.This paper addresses the challenge of accurately predicting FSO link quality in complex dynamic environments， with a specific focus on Q-factor prediction. Traditional physical models struggle to characterize nonlinear channel behavior， while existing neural network-based approaches often suffer from limited accuracy， slow convergence， and susceptibility to local optima. To overcome these challenges， a BP neural network integrated with a Newton-Raphson-based optimizer is proposed. By incorporating second-order optimization and a trap-avoidance mechanism， the NRBO-BP algorithm effectively adjusts weight update directions， suppresses ineffective training iterations， and enhances global search capability. Simulation results demonstrate that the proposed method significantly improves both prediction accuracy and training performance. The NRBO-BP algorithm provides a reliable and efficient solution for real-time FSO link quality assessment， offering strong support for adaptive performance optimization and stable system operation in complex dynamic environments.