A CNN–LSTM Hybrid Deep Learning Technique for Predicting Performance Degradation in SDH Networks
Abstract
The increasing complexity of Synchronous Digital Hierarchy (SDH) networks necessitates reliable forecasting of network performance indicators to support proactive performance monitoring. This study evaluates deep learning approaches for forecasting key performance indicators (KPIs) in SDH networks using real-world Network Management System (NMS) data from Uyo, Nigeria. The dataset comprised 52,128 observations collected at 15-minute intervals from January to June 2026 across three SDH links: UYO_EKET, UYO_ITU, and UYO_ORON. Five KPIs—Bit Error Rate (BER), Errored Seconds (ES), Severely Errored Seconds (SES), Background Block Errors (BBE), and Unavailable Seconds (UAS)—were examined. Data were divided chronologically into training (January–April), validation (May), and independent testing (June) periods. Four deep learning architectures—Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and CNN–LSTM—were evaluated for one-step-ahead (15-minute) forecasting using a 24-step (6-hour) historical sequence. Standardization was fitted exclusively on the training data to prevent information leakage. Performance was assessed using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). Results demonstrated that the CNN–LSTM model consistently produced the lowest MSE, RMSE, and MAE, while its prediction curves generally remained closer to the observed temporal behaviour than those of the individual RNN, CNN, and LSTM models. Across the KPIs, the models successfully captured the underlying temporal patterns but tended to smooth abrupt fluctuations and extreme observations, particularly in BBE, where the actual series exhibited substantial spikes. SES and UAS were predominantly concentrated around zero, resulting in relatively small absolute prediction errors despite some visible deviations in the plots. The consistency between the numerical metrics and graphical results indicates that the CNN–LSTM model provides more stable and accurate forecasting of SDH network KPIs. The findings demonstrate the potential of Hybrid Deep Learning architectures to enhance proactive network performance assessment through short-term forecasting of emerging degradation patterns, supporting the early identification of abnormal network conditions.