Shengfa Miao

Driven by non-stationary factors in real-world sensor-driven applications, time series streams from energy meters, traffic detectors, weather stations, and industrial monitors often exhibit complex patterns composed of long-term trends and multi-scale seasonal fluctuations. Accurately disentangling and modeling these heterogeneous components remains a fundamental challenge in long-term time series forecasting (LTSF). To address this issue, we propose STDPatch, a novel three-stream forecasting framework that combines structural decomposition with architecture specialization. First, we introduce an SG-Filter-Based seasonal–trend decomposition module that employs polynomial fitting to extract shape-preserving trends while reducing seasonal noise. Second, we design a trend decomposition module that further separates the trend component into ascending and descending segments to capture fine-grained evolutionary dynamics. Third, we propose a patch refactor module that adaptively aggregates adjacent patches according to structural similarity, thereby preserving temporal semantic continuity and reducing spurious correlations. Finally, we develop a three-stream architecture that leverages convolutional, linear, and Transformer branches to model seasonal patterns, smooth trends, and non-stationary sub-trends, respectively, with each branch built from efficient, channel-independent components. Extensive experiments on seven real-world sensor-derived benchmark datasets demonstrate that STDPatch consistently outperforms state-of-the-art methods for long-term time series forecasting.

Data pruning is a key technique for reducing training costs and improving model performance. However, most existing methods rely on fixed pruning rates or single metrics, which are largely designed for static training settings, making them unsuitable for incremental learning, where data distributions and model states change dynamically. To address this, we propose , a progressive and adaptive data pruning method for incremental learning, which dynamically evaluates sample difficulty and its difficulty changes during training to enable adaptive sample selection for each incremental learning task. Specifically, we propose two metrics, the instant difficulty score and the difficulty variation score. The former evaluates the learning difficulty of a sample, while the latter evaluates the variation in difficulty over a training interval. These two metrics are combined to guide pruning decisions. To prevent certain classes from being completely removed, we also introduce a class-balance retention mechanism. Experimental results show that outperforms existing data selection methods on CIFAR-100 and Tiny-ImageNet, generalizes across multiple incremental learning frameworks, and still maintains or exceeds the original accuracy even when training time is reduced by up to 52.90% compared to using the full dataset, demonstrating its effectiveness and practical value.