Sino-European Trade Flow Forecasting and Path Optimization: A GFIITL-DCL-MHA Framework Based on Deep Collaborative Learning and Transfer Learning

To address the challenges of macroeconomic volatility and "cold start" issues in Sino-European trade, this paper proposes the innovative GFIITL-DCL-MHA framework. This framework integrates a Multi-Head Attention (MHA) mechanism for feature collaboration and Particle Swarm Optimization (PSO) for global hyperparameter calibration. Empirical analysis conducted on trade data from 2021 to 2024 demonstrates that the framework achieves a superior Mean Absolute Percentage Error (MAPE) of 3.48%, significantly outperforming standard benchmark models. By leveraging the transfer learning module, the model achieves stable convergence in only 15 epochs, representing a 70.0% increase in training efficiency for data-sparse routes. Furthermore, with a high AUC of 0.92 for risk detection, the framework successfully optimized complex logistics paths, reducing the average transit time from 18.5 days to 14.2 days while lowering overall operational costs by 13.5%. These empirical results provide a robust, data-driven scientific basis for significantly enhancing the overall resilience and efficiency of the "Belt and Road" logistics network in an increasingly uncertain global environment.