AST-CLNet: An Image-Inspired Adaptive Soft Threshold and Dual-Path Spatiotemporal Modeling Method for Time Series Forecasting

Time series forecasting often faces challenges such as noise interference, complex feature relationships, and dynamic temporal dependencies. These issues prevent traditional models from effectively extracting key information, thus limiting prediction accuracy. To address these problems, inspired by soft-threshold techniques in image denoising, this study proposes a dual-path spatiotemporal forecasting model (AST-CLNet) integrated with adaptive soft-threshold denoising. The model improves time series forecasting through two core mechanisms: First, it introduces an image-inspired adaptive soft-threshold denoising module. Through residual connections, this module effectively filters out noise while preserving useful features, providing a cleaner data foundation for subsequent feature extraction. Second, it draws on the logic of “first optimizing local details, then modeling global semantics” in image processing. It first captures short-term fluctuations by extracting local features, then captures long-term dependencies by modeling global context. Finally, the model integrates local and global spatiotemporal features to make full use of multi-dimensional information. Experimental results show that AST-CLNet outperforms multiple benchmark models in various forecasting tasks, demonstrating its advantages in suppressing noise and capturing complex spatiotemporal relationships.