TSS-LCD: A Temporal–Spectral–Spatial-Guided Latent Conditional Diffusion Model for Spectrum Prediction Under Incomplete Observations

Accurate spectrum prediction provides the foresight needed for timely access decisions and proactive interference avoidance in diverse wireless scenarios. Since spectrum RSS exhibits inherent, coupled dependencies across the TSS dimensions reflecting underlying spectrum usage patterns, most existing methods extract TSS features from complete historical observations and map them to future RSS through simple regression structures. However, in practical deployments the historical observations are often incomplete, which corrupts or removes informative patterns and makes it difficult to joint capture TSS dependencies. Furthermore, spectrum RSS data contains fine-grained variations; direct feature-to-prediction tends to smooth these details, reducing prediction accuracy. To address these problems, we propose TSS-LCD, a two-stage network that jointly captures TSS dependencies and then uses them to guide diffusion process in latent space for spectrum prediction. Specifically, the TSS-CC stage employs three parallel self-attention branches and a cross-attention fusion module to extract and integrate TSS dependencies into a unified conditional representation. The LCD-SP stage then performs latent conditional diffusion, using this representation as conditioning at every denoising step to reconstruct detail-preserving future RSS data. Experiments on a real-world dataset show that TSS-LCD outperforms representative baselines, achieving lower errors and better recovery of fine-grained variations under incomplete historical observations.