TFT-MPIR: An end-to-end multi-period inventory replenishment strategy based on temporal fusion transformer

The inherent uncertainty associated with demand and vendor lead time significantly complicates replenishment strategies, which is challenge in the dynamic realm of e-commerce platforms. An end-to-end multi-period inventory replenishment strategy based on temporal fusion transformer (TFT-MPIR) is tailored for an integrated inventory replenishment decision-making process, and takes into account stochastic demand, vendor lead time, as well as linear transportation costs. TFT-MPIR which is fundamentally trained on an extensive amount of historical data, utilizes deep learning to directly calculate replenishment orders based on contextual and historical insights, deviating from the conventional two-step Predict-Then-Optimize (PTO) approach. The TFT-MPIR neural network framework designed with the concept of modularity enables an in-depth understanding and optimization of its structure and parameters. Specifically, the demand forecasting module utilizes temporal fusion transformer for advanced multi-quantile forecasting, generating comprehensive demand projections that significantly improve the accuracy of subsequent replenishment decisions. Numerical experiments incorporate authentic historical data from a prominent beverage supplier. Compared to the optimal solution (OPT) for inventory costs, TFT-MPIR exhibits a variance of 15.8%, markedly surpassing other integrated inventory strategies namely (t, R, S), PTO, and E2E-Multi-layer perceptron(E2E-MLP), which demonstrate divergences of 34.8%, 24.1%, and 22.3% respectively from OPT. Furthermore, TFT-MPIR framework achieves a cost reduction of 8.3% relative to the conventional PTO, and 19% in comparison to the (t, R, S). The robustness and scalability of the TFT-MPIR are substantiated through the adjustment of the ratio between unit stockout cost and unit transportation cost, coupled with sensitivity analysis.