Design of an improved model for demand forecasting and just in time inventory optimization using multi timescale encoding and graph-attention transformers

Traditionally, in supply chain management forecasts made by demand-replicating autoregressive models and classical machine learning approaches fail to consider all temporal and economic dependencies affecting demand variations in the process. Most of these standard transformers or any current deep learning-based models focus on their sequential dependencies but do not have explicit mechanisms to incorporate all these external economic factors, multi-scale temporal fluctuations, and, of course, the other interactions linked to the supply chain in process. A fused transformer-based demand forecasting model for Just-In-Time Inventory optimization is proposed. The proposed framework combines several advanced transformer architectures with reinforcement learning-based inventory control optimization techniques to solve optimization problems and make accurate forecasts. The proposed model shows significant improvement over existing models in terms of lower forecast errors and better synchronization of lead time. It also reduces occurrence of stockouts and overstocking thereby reducing inventory costs. The proposed model is a viable solution for improving the resilience and cost efficiency of supply chain operations.