Hybrid RK4-LSTM Model for Adaptive QoS Optimization in LTE eUTRAN: Deterministic vs Stochastic Approaches

BROU Pacôme, PANDRY Ghislain, ASKA Marcellin and OUMTANAGA Souleymane

Background and Objective: In the context of exponentially growing demand for mobile connectivity, the dynamic management of Quality of Service (QoS) in LTE networks represents a major challenge. The main issue lies in the difficulty of anticipating and adapting network resources in response to unpredictable traffic variations, especially during peak hours. This study aims to evaluate and compare the effectiveness of various simulation models RK4 (Fourth-order Runge-Kutta method) (deterministic and stochastic), LSTM (Long Short-Term Memory neural network) (deterministic and stochastic) and a hybrid RK4+LSTM model for forecasting and adaptively optimizing QoS over a full 24 hrs cycle that includes peak traffic periods (08:00-10:00 and 18:00-20:00). Materials and Methods: This study proposes a hybrid approach combining the numerical precision of the fourth-order Runge-Kutta (RK4) method with the predictive capacity of Long Short-Term Memory (LSTM) neural networks, explored under both deterministic and stochastic frameworks. Results: The hybrid RK4+LSTM approaches outperform standalone models by combining modeling accuracy and temporal anticipation. The deterministic approach offers an effective compromise between radio resource allocation (89.4%) and latency regulation (61.3 msec), while the stochastic approach provides the best peak anticipation (5-10 min), with an average allocation of 91.8%, bandwidth of 46.9 Mbps and an average delay of 79.3 msec during high-load periods, while maintaining low variability. In contrast, the LSTM-only model offers better anticipation (up to 20 min) but less stability, whereas the RK4 models (deterministic and stochastic), though effective in mathematical modeling, fail to capture sudden traffic fluctuations. Conclusion: The decisive value of mathematical-AI coupling in designing intelligent LTE networks capable of real-time resource adjustment and adaptive congestion anticipation.

View Fulltext Back