Deep Learning for Telecom Self-Optimized Networks: Benefits and Implications

Self-Optimized networks (SON) have emerged as a pivotal solution for telecom operators to automate their networks' implementation, configuration and resources optimization based on network's own intelligence. Among the challenges tackled by SON, traffic prediction stands out as a critical endeavor allowing dynamic and optimal resource allocation in the short term and giving a clearer visibility about network's future needs in terms of capacity and energy on the long run. However, most existing studies rely on highly complex models with low interpretability, resulting in inefficient solutions with substantial implementation and computational costs. This makes them unsuitable for real-world scenarios, where simplicity, transparency, and adaptability to dynamic conditions are critical for practical deployment. This study introduces an efficient traffic prediction approach that combines an innovative data partitioning strategy to capture spatial dependencies with Long Short-Term Memory (LSTM) networks to model temporal patterns. Leveraging real traffic data from a leading Moroccan telecom operator, the proposed model accurately forecasts future traffic patterns and their geographic distribution, achieving an absolute prediction error of less than 15 GB. These high-precision forecasts significantly improved network awareness, enabling the deployment of energy optimization strategies that reduced energy consumption across 1,100 base stations by an average of 11% per station.