<ul><li>Recent advances in Spatial-Temporal Graph Neural Network (STGNN) have led to significant progress in modeling spatial-temporal correlations for Multivariate Time Series (MTS) forecasting.</li><li>A new framework called Long-term Multivariate History Representation (LMHR) Enhanced STGNN is proposed to address the overlooked long-term spatial-temporal similarities and correlations in MTS, crucial for accurate forecasting.</li><li>LMHR incorporates a Long-term History Encoder (LHEncoder) to encode long-term history effectively, a Hierarchical Representation Retriever (HRetriever) for spatial information retrieval, and a Transformer-based Aggregator (TAggregator) for efficient fusion of retrieved contextual representations.</li><li>Experimental results show that LMHR outperforms typical STGNNs by 10.72% on average prediction horizons and state-of-the-art methods by 4.12% on real-world datasets, with consistent improvements in prediction accuracy for rapidly changing patterns.</li></ul>

Leveraging Multivariate Long-Term History Representation for Time Series Forecasting

Discover more