Optimization of Fiber Attenuation Prediction Based on GA-CNN-BiLSTM-Attention

Fiber optic communication is an important pillar of modern information transmission network, and fiber optic attenuation as the main factor of optical communication system performance degradation, the related prediction and maintenance research is of great practical significance, but the traditional prediction methods are still inefficient and have poor accuracy. In this paper, we propose a hybrid CNN-BiLSTM model based on the Attention mechanism (Attention) and optimize the model hyperparameters by combining with Genetic Algorithm (GA). By preprocessing the Optical Time Domain Reflectometer (OTDR) data, in the fiber attenuation prediction model, CNN is used to extract the local features of the time-series data, BiLSTM can effectively capture the long-term dependency of the time-series, and the Attention mechanism enhances the weight of the key timesteps, which improves the ability of the model to focus on the key features. The GA algorithm optimizes the hyperparameters of the model, which significantly improves the model's prediction accuracy and generalization ability. The experimental results show that the optimized GA-CNN-BiLSTM-Attention model outperforms the traditional method in the key metrics of prediction accuracy (98 %), RMSE (0.45) and coefficient of determination (0.92).