Link Performance Prediction of Power Fiber Optic Communication System Based on Attention Mechanism and Convolutional Neural Network Fusion

Abstract

In order to enhance the multi-objective optimization capability of power communication transmission networks, the author proposes an optimization method that integrates improved graph neural networks (GNNs) and genetic algorithms (GAs). The model integrates graph convolution and an attention mechanism to construct a multi-output prediction structure, achieving joint optimization of network reliability, transmission delay, and resource utilization. The test results on Institute of Electrical and Electronics Engineers (IEEE) 118 and 300 node systems show that this method significantly outperforms traditional Convolutional Neural Network (CNN) models in terms of network reliability (improved by 9.7%), latency (reduced by 24.7%), and resource utilization (improved by 11.5%). At the same time, the fusion model optimized the convergence algebra by 38% and increased the number of non-dominated solutions by 50%, demonstrating stronger solution space exploration ability and convergence efficiency. (1) Integrating a dynamic attention mechanism (graph attention module) with a residual graph convolution module to prioritize bottleneck links in power networks, unlike GraphCast’s fixed attention weights and (2) embedding GNN-derived features into GA initialization, addressing OpenDSS-GA’s reliance on random population generation. The research has verified the effectiveness and scalability of this method in largescale power communication networks, providing new ideas for optimizing complex networks. Index Terms—Attention mechanism, electric power communication transmission network, genetic algorithm, graph neural network, multi-objective optimization.

Cite this article as: Y. Zhang, Y. Liu, C. Wang and J. Dong, “Link performance prediction of power fiber optic communication system based on attention mechanism and convolutional neural network fusion,” Electrica, 26, 0162, 2026. doi: 10.5152/electrica.2026.25162.