Multivariate time series (MTS) forecasting is of critical importance in practical applications. Graph neural networks (GNNs) offer new insights for MTS forecasting, but traditional GNN methods rely on static graph structures, making it difficult to capture dynamic correlations and evolutionary patterns, and they also have limitations in the fusion of node and edge features. To address these challenges, this paper proposes a dynamic association multi-attribute fusion graph network (DyAMFG) for multivariate time series forecasting. The model first employs the association feature extraction and feature-driven edge learning mechanism to construct an adaptively evolving dynamic association graph, capturing the non-stationary patterns of node-edge co-evolution. Then, the complementary multi-feature encoders are designed to jointly model the neighbor aggregation, the neighbor co-occurrence, and the time dependence edge features, comprehensively covering dynamic changes and data trends. Finally, the adaptive fusion mechanism is used to break through the information barriers between node and edge features, achieving deep fusion across attribute features. Extensive experiments are conducted on five real-world datasets and the results validate that the DyAMFG model demonstrates outstanding prediction and generalization performance. Compared with other reported methods, the DyAMFG model achieves average improvements of 37.9%, 42.5%, and 11.7% in the RRSE metric across three datasets, and improves the RMSE metric by 25.8% and 6.90% on the remaining two datasets.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号