CBEC inventory optimization model design based on spatiotemporal attention and transformer architecture

Area of Science:

• Supply Chain Management
• Artificial Intelligence
• Data Science

Background:

• Cross-border enterprises face challenges with inaccurate long-term inventory predictions.
• Existing models struggle to capture complex spatiotemporal dynamics in inventory data.

Purpose of the Study:

• To develop an advanced cross-border inventory prediction model.
• To improve the accuracy and reliability of long-term inventory forecasting.

Main Methods:

• Utilized a spatiotemporal perception Transformer incorporating a temporal aware self-attention mechanism.
• Employed multi-scale diffusion convolution to simulate inventory data diffusion across road networks.
• Implemented a multi-dimensional feature fusion module for adaptive spatiotemporal feature integration.

Main Results:

• The proposed model demonstrated superior performance compared to the ASTGNN model.
• Achieved reductions in Mean Absolute Error (MAE) by 7.6%, Mean Absolute Percentage Error (MAPE) by 4.2%, and Root Mean Squared Error (RMSE) by 1.1%.

Conclusions:

• The spatiotemporal perception Transformer model offers a significant advancement in cross-border inventory prediction.
• The model effectively captures global and local spatiotemporal features for enhanced forecasting accuracy.