This study aims to construct and optimize a prediction model for agricultural product consumer preferences. Utilizing a deep learning framework combining LSTM with attention mechanism and integrating model ensemble strategies significantly improves prediction performance. Results show that the model outperforms traditional methods across various evaluation metrics, achieving a MAPE reduction to 6.94% and an accuracy rate of 85.1%. The model demonstrates good stability and generalization ability in a 6-month rolling forecast. By introducing periodic neural network components, further optimization for seasonal agricultural product prediction is achieved. These findings provide reliable data support for agricultural production planning and market decisions, potentially driving intelligent development in the agricultural sector.