Convolutional neural network and long short-term memory forecasting and variational autoencoder anomaly detection in 4G cellular networks

Abstract

Spectrum monitoring in cellular networks with limited resources remains predominantly reactive because capacity planning and failure detection occur after network congestion, while spectrum monitoring data are rarely available for reproducible research. Existing studies address short-term traffic forecasting and anomaly detection as separate tasks and largely depend on densely sampled operator telemetry data that are inaccessible to academia and regulators in emerging markets. This study integrates both tasks into a unified deep-learning pipeline for a 4G cellular environment. The Zimbabwe spectrum dataset contains 315,247 hourly measurements collected from 13 cell sites, three operators, and five frequency bands, annotated with four operator-defined anomaly classes representing 2.03% of all measurements. A hybrid one-dimensional (1-D) convolutional neural network-long short-term memory (CNN–LSTM) model uses a 72-hour traffic window to forecast the next six hours of aggregate network traffic, while a variational autoencoder (VAE) trained exclusively on normal records detects anomalies when reconstruction error exceeds the 99th-percentile validation threshold. On the held-out test set, the model achieved a mean absolute error (MAE) of 158.97 GB, root mean square error (RMSE) of 230.41 GB, and mean absolute percentage error (MAPE) of 52.93%, outperforming a seasonal-naive baseline (MAE=216.43 GB, MAPE=66.77%, p<0.001, Diebold-Mariano test). The study contributes a publicly available localized synthetic dataset, a reproducible end-to-end forecasting and anomaly detection baseline, and a per-class anomaly analysis.