谢骋辉 题为 “STRFormer: Shared Temporal Reference Transformer for Long-Horizon Time Series Forecasting” 的论文已被《Expert Systems with Applications》接受发表。该论文摘要如下：

Time-series forecasting is widely applied in traffic systems, energy management and space telemetry analysis, where reliable long-horizon prediction is essential. In multivariate long-horizon forecasting, performance depends not only on capturing temporal dynamics within each variable but also on modeling stable dependencies across variables. Although existing method has explored modeling global inter-variable correlations within the attention mechanism, it only reconstructs the query. This neglect of reconstructing the key and value leads to inaccurate similarity estimation in the attention computation. To address this issue, we propose the Shared Temporal Reference Transformer (STRFormer). STRFormer introduces a shared temporal reference mechanism that aligns query, key, and value representations within the attention module, establishing a unified temporal coordinate system for cross-variable interactions. This design encourages more coherent channel-wise attention patterns without modifying the overall attention structure. In addition, STRFormer incorporates a Multi-scale Refinement module that jointly models global temporal structure and local dynamics, forming a multi-scale architecture for long-horizon forecasting. Extensive experiments on thirteen public datasets from diverse domains demonstrate consistent improvements over competitive baselines, with particularly strong performance on high-dimensional traffic benchmarks.