Globally, the telecommunications sector has undergone significant transformations. Driven by rapid technological progress, increasing teledensity, and evolving consumer demographics, this industry is at a crucial juncture. These transformations are characterized by frequent deterioration in the quality of service (QoS) offered by telecommunications service providers, which results in poor experiences for consumers and inefficient network resources. To address this gap, this study investigates the application of deep learning techniques for long-term evolution (LTE) traffic demand prediction using real-world, anonymized cellular data obtained from a sub-Saharan African telecommunications service provider. Precisely, three long short-term memory (LSTM) neural network architectures, vanilla LSTM, stacked LSTM, and bidirectional LSTM, are developed and evaluated under different loss functions, with the Adam optimizer employed during training. The study demonstrates the significance of choosing an appropriate loss function based on the inherent characteristics of the problem, data distribution, and model constraints to enhance each prediction. The predictions were evaluated using the coefficient of determination R2 metric. The mean squared error loss function was determined to be the most effective for LTE traffic demand prediction task, achieving an accuracy of 86. 44% with the vanilla LSTM, while the bidirectional LSTM reached 85. 43% accuracy by using the SmoothL1 loss function during learning. The findings establish that increased architectural complexity does not necessarily translate into improved performance for LTE traffic demand prediction in data-constrained environments. Consequently, the vanilla LSTM is recommended as an effective and computationally efficient solution for traffic prediction and QoS-aware resource management in LTE networks within sub-Saharan Africa.
Efuwape et al. (Wed,) studied this question.