Multi-Model Fusion Demand Forecasting Framework Based on Attention Mechanism

Abstract

The accuracy of demand forecasting is critical for supply chain management and strategic business decisions. However, as data volumes grow and demand patterns become increasingly complex, traditional forecasting methods encounter significant challenges in processing intricate multi-dimensional data and achieving a satisfactory predictive accuracy. To address these challenges, this paper proposed an end-to-end multi-model demand forecasting framework based on attention mechanisms. The framework employs a dual attention mechanism to dynamically extract features from both the temporal and product dimensions, while integrating conditional information captured through convolutional neural networks, thereby enhancing its ability to model complex demand patterns. Additionally, a channel attention mechanism is introduced to perform the weighted fusion of outputs from multiple predictive models, thereby overcoming the limitations of single-model approaches and improving adaptability to varying demand patterns across diverse scenarios. The experimental results demonstrate that the proposed method outperforms conventional approaches across several evaluation metrics, achieving a 42% reduction in Mean Squared Error (MSE) compared to the baseline model. This notable improvement enhances both the accuracy and stability of demand forecasting. The framework offers valuable insights for addressing large-scale and complex demand patterns, providing guidance for precise decision-making and resource optimization within supply chain management. Future research will concentrate on further enhancing the model’s generalization capability to manage missing data and demand fluctuations. Additionally, efforts will focus on integrating diverse heterogeneous data sources to assess its performance in various practical scenarios, ultimately improving the model’s accuracy and flexibility.