Probabilistic Interference Prediction for Dynamic 6G In-X Sub-Networks

In-X Subnetworks (SNs) face significant challenges in meeting extremely heterogeneous requirements due to interference resulting from dynamic mobility, hyper-dense deployment, and limited available channel bandwidth. To address these challenges, we introduce novel probabilistic interference prediction techniques that enable proactive interference management by allocating resources based on predicted interference, thereby preventing performance degradation in SNs. However, interference prediction is challenging due to complex, dynamic interference associated with randomness in the propagation environment, mobility, and traffic patterns. We propose and evaluate two categories of probabilistic predictors to model the tail statistics of interference: a Bayesian framework-based Variational Inference Sparse Gaussian Process Regression (VISPGPR) and a Quantile Bidirectional Long Short-Term Memory (QBiLSTM) predictor for direct quantile estimation. Additionally, we introduce a Matérn kernel-based Student-t process to effectively capture interference dynamics characterized by outliers and rapid fluctuations. This approach outperforms traditional methods, including VISPGPR with the commonly used exponential square kernel and Gaussian prior, by better modeling heavy-tailed interference distributions. However, model mismatch has been identified as a major bottleneck for optimal performance. To address this, we propose an Attention-based modified QBiLSTM (Atten-MQBiLSTM), which further enhances performance by leveraging temporal correlations without making assumptions about the underlying distribution of target variables-an aspect not captured by VISPGPR. The effectiveness of the proposed predictors is evaluated using a spatially consistent 3GPP channel model incorporating realistic mobility and two distinct extreme traffic models. Simulation results show that the proposed predictors outperform the baseline method in terms of coverage probability, with improvements ranging from 22–33% and 4–13% over VISPGPR for Bernoulli random and deterministic traffic, respectively. This helps achieve target reliability close to Genie resource allocation (RA) in both extreme traffic modeling scenarios while ensuring scalability within the predictive interference management framework.