Scholarlink

<p>Time series forecasting plays a vital role in domains such as energy consumption estimation, weather analysis, and traffic flow prediction, where accurate predictions are crucial for effective planning and decision-making. However, current methods often struggle to model complex interdependencies among multiple variables and localized dynamics in long-range sequences, which limits forecasting performance. To overcome these limitations, this paper introduces the Multi-view Graph Representing Interactive Learning Network (MGRILN). The model leverages graph convolutional networks to build topological structures from three complementary perspectives: temporal, dimensional, and cross-segmental. It uses differential operations to capture inter-dimensional trend shifts and constructs cross-segment graphs to represent local volatility correlations. A multi-scale interaction module is also integrated to model dependencies across different temporal resolutions. Comprehensive experiments on five real-world datasets show that MGRILN reduces MAE and MSE by an average of 8.7% and 9.2%, respectively, outperforming state-of-the-art benchmarks. For example, on the Weather dataset with a prediction horizon of 720 steps, it achieves an MAE of 0.334. These results demonstrate the model’s strong forecasting capability and practical value.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-view graph representing interactive learning network for time series forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月18日

查看全文
  • Guanshu Wang,
  • Limin Liu,
  • Bin Wu

摘要

Time series forecasting plays a vital role in domains such as energy consumption estimation, weather analysis, and traffic flow prediction, where accurate predictions are crucial for effective planning and decision-making. However, current methods often struggle to model complex interdependencies among multiple variables and localized dynamics in long-range sequences, which limits forecasting performance. To overcome these limitations, this paper introduces the Multi-view Graph Representing Interactive Learning Network (MGRILN). The model leverages graph convolutional networks to build topological structures from three complementary perspectives: temporal, dimensional, and cross-segmental. It uses differential operations to capture inter-dimensional trend shifts and constructs cross-segment graphs to represent local volatility correlations. A multi-scale interaction module is also integrated to model dependencies across different temporal resolutions. Comprehensive experiments on five real-world datasets show that MGRILN reduces MAE and MSE by an average of 8.7% and 9.2%, respectively, outperforming state-of-the-art benchmarks. For example, on the Weather dataset with a prediction horizon of 720 steps, it achieves an MAE of 0.334. These results demonstrate the model’s strong forecasting capability and practical value.