Scholarlink

<p>As global supply chains become increasingly susceptible to these risks, demand forecasting models must be both accurate and efficient across temporal and spatial dimensions. Standard models, such as ARIMA, RNNs, and Transformers, often underperform when handling nonlinear dynamics, long-sequence dependencies, and heterogeneous data. To overcome these challenges, we propose a multi-scale spatiotemporal forecasting framework based on the Mamba state space backbone, enhanced with dynamic graph attention and event-aware embeddings. This unique method incorporates trade flows, spatial connections from OSM, news from Reuters/GDELT, and economic indicators, using seasonal decomposition, spatial graph embeddings, and contrastive learning to model complex interactions across scales. The experimental comparison with ARIMA, Transformer, and TSMamba baselines demonstrates that the proposed model achieves more than 90% accuracy consistently across different horizons, improving MAE/RMSE by 25% while reducing FLOPs by 60% compared to Transformers. Our approach has been established as more accurate and more efficient, hence providing a realistic solution to resilient global supply chain management.</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-scale global supply chain spatiotemporal demand forecasting based on Mamba SSM state space backbone

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月22日

查看全文
  • Lin Gao,
  • Kebin Lu

摘要

As global supply chains become increasingly susceptible to these risks, demand forecasting models must be both accurate and efficient across temporal and spatial dimensions. Standard models, such as ARIMA, RNNs, and Transformers, often underperform when handling nonlinear dynamics, long-sequence dependencies, and heterogeneous data. To overcome these challenges, we propose a multi-scale spatiotemporal forecasting framework based on the Mamba state space backbone, enhanced with dynamic graph attention and event-aware embeddings. This unique method incorporates trade flows, spatial connections from OSM, news from Reuters/GDELT, and economic indicators, using seasonal decomposition, spatial graph embeddings, and contrastive learning to model complex interactions across scales. The experimental comparison with ARIMA, Transformer, and TSMamba baselines demonstrates that the proposed model achieves more than 90% accuracy consistently across different horizons, improving MAE/RMSE by 25% while reducing FLOPs by 60% compared to Transformers. Our approach has been established as more accurate and more efficient, hence providing a realistic solution to resilient global supply chain management.