LXformer: A Long-Term Time Series Forecasting Model Based on Multi-Granularity Feature Extraction for Edge Devices

Multivariate time series forecasting is a fundamental research problem in the Internet of Things (IoT), as it provides critical decision support and underpins the accuracy and reliability of downstream intelligent systems. Existing approaches commonly rely on sequence decomposition strategies that separate time series into trend, seasonal, and residual components. However, limited attention has been paid to how temporal information can be represented and integrated across different granularities. Inspired by the human reading process, in which information is progressively understood at the word, sentence, and paragraph levels, we propose LXformer, a novel forecasting framework that captures multiscale temporal representations. After segmenting multivariate time series into patches, LXformer integrates information at multiple granularities by modeling intra-patch features, inter-patch dependencies, and inter-sequence relationships to accomplish the forecasting task. Specifically, multiple one-dimensional convolutional branches are employed to extract fine-grained local patterns within each patch from diverse perspectives. In addition, agent attention is introduced to facilitate effective interactions across patches and channels, enabling the modeling of coarser-grained temporal dependencies. The combination of one-dimensional convolutions and linear-complexity attention mechanisms ensures that LXformer maintains overall linear computational complexity. Extensive experiments conducted on nine large-scale real-world datasets demonstrate that LXformer consistently achieves lower forecasting errors while delivering faster inference speed and reduced memory consumption. These advantages make LXformer particularly suitable for deployment on edge devices with limited computational resources but high accuracy requirements.