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

## TL;DR

This paper introduces a new model for predicting cross-border inventory using spatiotemporal attention and Transformer architecture to improve prediction accuracy.

## Contribution

The novel contribution is an improved temporal aware self-attention mechanism and multi-scale diffusion convolution for cross-border inventory prediction.

## Key findings

- The proposed model outperforms ASTGNN in MAE, MAPE, and RMSE by 7.6%, 4.2%, and 1.1%, respectively.
- The model captures global spatiotemporal features and local spatial features using a spatiotemporal correlation matrix and multi-scale diffusion convolution.

## Abstract

To solve the problem of inaccurate long-term prediction of cross-border inventory encountered by cross-border enterprises in their experience, this paper proposes a cross-border inventory prediction model based on a spatiotemporal perception Transformer. Specifically, firstly, an improved temporal aware self-attention mechanism is adopted to mine potential temporal trends and spatial heterogeneity features in cross-border inventory, and an accurate spatiotemporal correlation matrix is established to obtain global spatiotemporal features. Secondly, we simulate the multi-level diffusion process of inventory data in the road network using multi-scale diffusion convolution, which captures the local spatial features of nodes across multiple neighborhood ranges. Finally, a multi-dimensional feature fusion module is used to adaptively fuse the captured spatiotemporal features and output prediction results. The experimental results show that compared with the ASTGNN model with the highest prediction accuracy, the method proposed in this paper performs better in MAE, MAPE, and RMSE, which are reduced by 7.6%, 4.2%, and 1.1%, respectively.