Improving Cellular Traffic Prediction with Temporal Embeddings: A Time2Vec-LSTM Approach

Abstract

The network traffic prediction has to be reliable for better resource allocation and congestion management in presentday telecommunications. In this paper, a novel hybrid Time2Vec-enhanced LSTM method is presented for somewhat more accurate traffic volume forecasting. The model exploits both historical traffic behavior and temporal features enriched by Time2Vec, such as hour and day, to represent the linear or periodic dependencies embedded in cellular traffic. Unlike traditional static time encodings or raw time features, the learnable Time2Vec embeddings enable the model to better capture daily and hourly fluctuations in network traffic. The study carried out experiments with a real-world dataset that had been collected from an LTE base station located in Kandahar Province of Afghanistan, with hourly uplink, downlink, and total traffic volumes recorded for 30 days. Performance was measured in terms of the Root Mean Square Error (RMSE) and coefficient of determination (R 2 ). The results show that the proposed Time2Vec-enhanced LSTM consistently outperforms Deep Learning (DL), statistical, and Machine Learning (ML) models across all traffic types. The learnable temporal embeddings are useful as they allow greater accuracy and better capture of trends. Ablation studies have supported that forecasting is far better with adaptive Time2Vec encoding than with models without or with a fixed-time feature, suggesting that learnable temporal features are essential for precise and robust cellular traffic prediction.