Accurately predicting commodity demand in international trade is crucial for making import decisions. This paper proposes a deep learning-based demand forecasting model that integrates Long Short-Term Memory (LSTM) networks, attention mechanisms, and a hierarchical prediction framework. It effectively captures temporal dependencies, identifies key features, and predicts total and segmented demands in stages. Tested on a large-scale dataset covering 205 countries and regions and 1,268 commodities, the model demonstrates outstanding forecasting capabilities with a comprehensive Mean Absolute Percentage Error (MAPE) of 9.7%. Particularly, it achieves high accuracy of 6.8% and 7.2% for staple agricultural products like wheat and soybeans, significantly outperforming traditional machine learning models. This model provides robust technical support for enterprises and governments to formulate precise international trade decisions.