PAD-STT: A Pre-Denoising Adaptive Decomposition Spatial–Temporal Transformer for Cellular Traffic Prediction

Abstract

Accurate prediction of cellular traffic is essential for enabling intelligent networks. However, this requires effectively capturing the dynamic variations within the traffic. Thus, we propose a Pre-Denoising Adaptive Decomposition Spatial-Temporal Transformer (PAD-STT). First, to mitigate noise interference in sequence decomposition, we apply a Savitzky–Golay filter for pre-denoising. Second, a dual-correlation mechanism is designed based on cross-correlation theory, enabling PAD-STT to perform synchronized modeling by jointly learning spatial and temporal relationships rather than in a sequential manner. Finally, we develop an adaptive decomposition to replace the original moving average, aiming to effectively capture dynamic variations. Experiments on real-world datasets demonstrate that PAD-STT outperforms representative state-of-the-art methods.