3D convolutional Dimension Preserved decomposition model integrated with Fourier Transform for traffic data prediction

Purpose The purpose of this study is to overcome the limitations of traditional deep learning architectures in processing missing and noisy spatio-temporal streaming data. It aims to develop a robust spatio-temporal forecasting framework that can directly utilize incomplete inputs to achieve accurate predictions, thereby preventing catastrophic performance degradation and gradient collapse. Design/methodology/approach An innovative framework, termed FTI3DCDP, is proposed for traffic forecasting with incomplete data without requiring prior imputation. It establishes a dual-domain learning mechanism, where Fourier Transform extracts robust periodic patterns from incomplete signals in the frequency domain, while Dimension Preserved (DP) Decomposition captures low-rank spatial structures. A 3D Convolutional Neural Network (3D-CNN) adaptively integrates these complementary features for end-to-end spatio-temporal forecasting. Extensive experiments on five real-world traffic datasets under varying missing rates demonstrate the effectiveness and applicability of FTI3DCDP across diverse traffic scenarios. Findings The results demonstrate that the proposed approach not only maintains forecasting stability under high sparsity but also consistently outperforms state-of-the-art baseline methods. This offers a highly resilient and deployable tool for supporting edge-intelligence solutions in modern transportation decision-support systems. Originality/value The study introduces a novel dual-domain spatio-temporal forecasting model differing from traditional prediction methods. The work systematically integrates Fourier Transform with Dimension Preserved Tensor Decomposition and 3D-CNN, and demonstrates substantial practical value through rigorous validation on multiple large-scale traffic datasets under severe data missing conditions.