Improved Multi-Step, Multi-Variate, and Spatiotemporal 5G Data Usage Forecasting Without Deploying Data Imputation Techniques

Accurately predicting 5G network usage is essential for efficient network resource allocation and high-quality service delivery. However, dynamic usage patterns, nonstationary usage distribution, and the presence of a high amount of missing data make forecasting future network usage statistics challenging. To address these challenges, this paper proposes a deep learning based time series forecasting approach that handles multi-step, multi-variate, and spatiotemporal time series forecasting problems. The proposed approach scans the data for continuity of time-steps and identifies an optimal sequence length in order to address the issue of missing records in supervisory data. It uses separate BiLSTM layers to capture the impact of temporal 5G data usage changes in adjacent networks. Next, it utilizes the adjacent cell's impact captured in permutationally invariant ordering for the central/targeted cell. The deep learning techniques used to capture the impact of changes in neighboring cells on the target cell and for multi-step and multivariate forecasting are: Self-attention, a RepeatVector, a BiLSTM network, and a Time distributed dense layer. The proposed approach outperforms existing state-of-the-art methods on the Milan dataset and has the potential to improve forecast accuracy in various applications, including network traffic forecasting.