A slice-aware deep learning framework for the joint prediction of traffic demand and multiple KPIs within a simulation-driven environment and shows that deep learning models more effectively capture nonlinear slice-level dynamics compared with traditional forecasting approaches.
Abstract
Accurate prediction of traffic demand and slice-level key performance indicators (KPIs) is essential for enabling proactive resource management in next-generation radio access networks. However, most existing studies focus on aggregate traffic modeling and provide limited insight into slice-level dynamics under varying mobility and traffic conditions. This paper proposes a slice-aware deep learning framework for the joint prediction of traffic demand and multiple KPIs within a simulation-driven environment. An ns-3-generated multivariate time-series dataset is constructed to capture traffic demand, throughput, goodput, latency, and packet-level statistics across eMBB, URLLC, and mMTC slices under heterogeneous mobility patterns, stochastic UE populations, varying traffic loads, and adaptive radio configurations. This enables a controllable and reproducible evaluation of slice-level traffic and KPI dynamics under diverse service conditions. LSTM, GRU, CNN, and a traditional linear regression baseline are comparatively evaluated under a unified preprocessing and time-series validation framework. Experimental results demonstrate that CNN consistently achieves lower prediction errors across most KPIs, while recurrent models provide competitive performance for smoother traffic patterns. The results further show that deep learning models more effectively capture nonlinear slice-level dynamics compared with traditional forecasting approaches. Latency prediction remains the most challenging task due to its sensitivity to rapid traffic fluctuations and varying network conditions. Beyond prediction, the proposed framework provides a foundation for learning slice-level network dynamics and supporting future digital twin–oriented network representations and closed-loop optimization in 6G-ready wireless systems.
3GPP Access Traffic Steering, Switching, and Splitting (ATSSS) enables traffic to be distributed across heterogeneous 3GPP and non-3GPP access networks to improve performance, reliability, and resilience. ATSSS can use multipath transport protocols such as Multipath QUIC (MP-QUIC), where packet scheduling plays a central role in determining latency and resource utilization for delay-sensitive applications. Many existing MP-QUIC scheduling policies rely on instantaneous path measurements or fixed rules rather than forecasts of future application demand. In multi-flow scenarios, such decisions can lead either to contention on the preferred low-latency path and transient latency inflation for priority traffic or to overly conservative use of available capacity. This paper proposes iScavenger, a predictive, machine-learning-based multi-flow scheduling policy for ATSSS environments. iScavenger employs a Long Short-Term Memory (LSTM) model to forecast near-future bandwidth demand for delay-sensitive Sticky traffic. Based on this prediction, background packets are admitted to the preferred low-latency path only when sufficient residual capacity is expected to remain; otherwise, they are steered to the alternative path. The policy is implemented within the Monty MP-QUIC framework and evaluated in a controlled Mininet testbed using traffic traces from the online game League of Legends, with fixed and variable path capacities and controlled jitter and packet loss. The results indicate that, under the evaluated conditions, iScavenger provides configurable operating points in the latency–utilization trade-off, limiting additional Sticky-flow RTT while achieving higher background throughput than conservative baseline policies. These findings highlight the potential of short-term traffic-demand prediction for proactive contention management in ATSSS-enabled multi-access networks.
Shah M. Emad Uddin, Karl-Johan Grinnemo, Arunselvan Ramaswamy et al.· IEEE Open Journal of the Com...· 0 citations
The study concludes that rule-based approaches remain highly valuable for mission-critical 5G slicing applications and recommends their integration with adaptive techniques to enhance scalability and performance.
R. Paper, Zayyanu Yunusa, Usman Haruna· International Journal of Eme...· 0 citations
The rapid growth of multimedia streaming poses critical challenges, including bursty traffic and congestion, leading to playback delays. The existing separate prediction and control mechanisms for multimedia traffic scheduling, which are based on software-defined networks (SDN), are unable to proactively manage bursty traffic under uncertain conditions. This limitation is particularly evident in SDN-enabled backbone and multimedia-aware access networks, which typically assume centralized control and stable topologies. They lack integration of traffic prediction, traffic shaping, and real-time perception scheduling through reinforcement learning, resulting in low efficiency when exploring multiple paths in dynamic networks. To address this challenge, we propose PPO-MS (Proximal Policy Optimization-based Multimedia Scheduler), an SDN-based multimedia traffic scheduling algorithm integrating three key innovations: 1) A novel LSTM+HTB synergy where LSTM’s confidence intervals dynamically adjust HTB (Hierarchical Token Bucket) shaping parameters, enabling adaptive rate control under prediction uncertainty and overcoming the limitations of static LSTM+HTB hybrids; 2) A Deep Reinforcement Learning (DRL)-optimized path pruning method that reduces state and action spaces by generating a constrained set of $k$ disjoint candidate paths via an improved redundant tree algorithm. Unlike traditional multi-path schemes, this method tightly couples path preselection with the RL decision loop for adaptive, context-aware routing; 3) Generalized Advantage Estimation (GAE)-accelerated PPO for stable convergence in dynamic environments. In contrast to prior works (e.g., LSTM+RL for QoE or standalone tree algorithms), PPO-MS uniquely unifies these modules through confidence-aware traffic shaping and hierarchical decision-making, validated via comparative experiments. Results demonstrate that PPO-MS, through the synergistic integration of confidence-aware traffic shaping and DRL-optimized path pruning, significantly outperforms decoupled baselines. In particular, via isolation studies against simpler alternatives (e.g., mean-prediction and fixed-margin shaping), the confidence-aware shaping mechanism is validated to be superior under bursty traffic conditions. Overall, PPO-MS reduces end-to-end latency by 17.3% and packet loss by 32.4% while achieving 24.4% better load balancing during traffic bursts.
Accurate network traffic forecasting is fundamental to Quality of Service enforcement, proactive congestion control, and dynamic resource allocation in modern backbone and software-defined networks. However, existing approaches often lack adaptability to non-stationary traffic patterns and fail to provide a consistent comparative evaluation across diverse models under unified experimental conditions. This paper presents an adaptive, data-driven framework that integrates Bidirectional LSTM (Bi-LSTM), LSTM, Gated Recurrent Units (GRU), Random Forest, XGBoost, Support Vector Regression (SVR), and classical ARIMA regressors for short- and medium-term traffic forecasting. The proposed architecture couples multi-scale temporal feature extraction with a feedback-driven online retraining loop, enabling continuous adaptation to distributional shifts. Extensive experiments are conducted on two publicly available datasets CAIDA Equinix backbone traces and the MAWI traffic archive comprising over 72 hours of flow-level measurements at one-minute resolution. Bi-LSTM achieves the lowest RMSE of 0.0287 Gbps and the highest $R^{2}=0.9714$, outperforming ARIMA by $\mathbf{7 5 . 2 \%}$ and vanilla LSTM by $\mathbf{8 . 0 \%}$. All results are confirmed via paired Diebold-Mariano (DM) tests and Student’s t-tests $(p \lt 0.01)$. System inference latency of 2.3 ms per batch satisfies real-time SDN control-plane requirements. Code and preprocessing scripts will be made publicly available to ensure full reproducibility.
E. Chithra, G. C. Bharathi, Sankara Rao et al.· International Conference on...· 0 citations
Point-to-multipoint (P2MP) coherent optical architectures that use digital subcarrier multiplexing support efficient aggregation in metro-access networks. Proactive provisioning of hub capacity requires accurate per-spoke demand forecasts. However, spoke nodes carry heterogeneous traffic profiles such as business, residential, and mixed, with distinct diurnal patterns and forecast difficulty. We examine whether a single multivariate Long Short-Term Memory (LSTM) network can capture these profile-specific dynamics without explicit labels, and quantify how the resulting per-profile forecast accuracy propagates to operational P2MP resource provisioning metrics. The forecasts feed a greedy reconfiguration-aware heuristic algorithm, which jointly decides light-tree assignment at each control epoch while accounting for rerouting, resizing, and point-to-point forwarding penalties. Evaluation on a 30-node TID-derived metro topology with 300 spokes with different OSNR budgets shows that the joint LSTM reduces per-profile mean absolute error by 13-22% over a seasonal baseline, with the largest gains on business spokes. LSTM-driven provisioning achieves 13-15% lower total operational cost than seasonal-driven provisioning and 29-30% savings over static peak allocation, while maintaining demand violations below 1.2%. The LSTM reduces reconfiguration churn by 12–26% per profile over seasonal and lagged baselines, largest on mixed spokes whose composite weekday/weekend pattern is hardest for seasonal forecasts to track.
P. Soumplis, Konstantinos Christodoulopoulos, K. Yiannopoulos et al.· International Symposium on C...· 0 citations
In dynamic Vehicular Edge Computing (VEC) environments, rapidly changing vehicle mobility and traffic lead to fluctuating edge resource demands, challenging task offloading and scheduling. Accurate prediction of future traffic flow and traffic-state-derived workload representations is thus crucial for proactive resource management. To address the limitations of existing methods in short-term dynamic characterization, complex spatial interaction modeling, and heterogeneous target prediction, this paper proposes a Spatio-Temporal Graph Attention Network (TA-STGAT). The proposed model constructs multi-dimensional RSU-level state sequences from simulated trajectories generated on a real-world road network and separately forecasts vehicle flow within RSU coverage areas and the associated traffic-state-derived workload representation under a unified spatio-temporal modeling framework. By integrating gated dilated temporal convolutions with a topology-constrained multi-head graph attention mechanism, the model captures multi-scale temporal dependencies and nonlinear spatial correlations. Experimental results show that, compared with the best-performing baseline in terms of RMSE for each forecasting task, TA-STGAT reduces RMSE by 10.89% and 11.29% in workload-representation prediction and traffic flow prediction, respectively, demonstrating its effectiveness for short-term edge-state forecasting.
