Supply chain demand forecasting based on multi-time scale data fusion network

Demand forecasting is crucial in supply chain management. However, the supply chain has grown increasingly complex and uncertain due to the expansion of mass manufacturing. Presently, prevailing supply chain demand forecasting methods primarily rely on traditional statistical models or single time scale models, resulting in inadequate forecasting performance. This paper proposes a novel supply chain demand forecasting model that employs multi-time scale data fusion to enhance forecasting performance. Initially, three parallel temporal convolutional networks are constructed to extract feature information from three different time scales of data. Following that, a feature fusion method is devised utilizing autoencoders and attention mechanism. The method reconstructs data through autoencoders, unifying the time dimension of features while enhancing them. It then adaptively calculates weights for different time scales using an attention mechanism and combines the reconstructed data with corresponding weights to obtain features containing information from various time scales. Subsequently, the proposed model further maps the fused data using temporal convolutional networks to obtain the final prediction output. Finally, this paper validates the model using data provided by a major home appliance manufacturer and compares it with various advanced time series forecasting models, demonstrating the superiority of the proposed model. The ablation experiments assess the effectiveness and necessity of the multi-time scale model input and fusion module on forecast results, confirming the superiority and necessity of the multi-time scale data input and fusion module.