Realizing the 6G vision of universal coverage necessitates reliable dynamic spectrum access to extend network reach into underserved regions. However, spectrum occupancy prediction in Cognitive Radio Networks (CRNs) is fundamentally compromised by additive white Gaussian noise (AWGN) which is an irreducible impairment. This paper presents a novel energy difference-based Long Short-Term Memory (LSTM) framework for robust multi-step spectrum occupancy prediction under AWGN. We introduce two reconstruction mechanisms, global scaling (ED-RM1) and variance-based (ED-RM2) conditioning to explicitly mitigate noise accumulation and training-testing SNR mismatch, issues often overlooked in existing energy and binary prediction methods. Unlike conventional predictors, the proposed framework maintains stable error performance across diverse occupancy dynamics and significant SNR gaps. Extensive simulations demonstrate that ED-RM2 achieves superior generalization and robustness, offering a resilient sensing solution essential for open and accessible 6G architectures.

CHATHURANGANI ALAHAKOON Dinushika;  KANDEEPAN Sithamparanathan;  YU Xinghuo;  KE Wang;  BALDINI Gianmarco; 

2026-06-01

IEEE

JRC145397

2836-4090 (online),   

https://ieeexplore.ieee.org/abstract/document/11517669,    https://publications.jrc.ec.europa.eu/repository/handle/JRC145397,   

10.1109/CSPA68262.2026.11517669 (online),