RDFT: A Resilience-Aware Deep Learning Framework for Probabilistic Demand Forecasting in Supply Chain Management

Good demand forecasts help with inventory control, replenishment and distribution planning. However, retail sales can change quickly due to promotions, seasonal effects, store-level differences and unexpected events. We propose the resilient demand fusion transformer (RDFT), which combines past sales, known future information, static attributes and resilience-related context in a multi-horizon framework. Adaptive variable selection is used to identify the most relevant demand drivers and a bidirectional GRU to capture short-term temporal movement. Resilience-aware attention retrieves informative earlier observations under the current operating conditions, and a gated fusion mechanism combines the resulting local and contextual representations. Then a quantile decoder produces 14-day forecasts and uncertainty ranges. We evaluated the framework on the Kaggle Store Sales dataset under a chronological split against random forest, XGBoost, LSTM and GRU. The RDFT achieved an MAE of 0.481, RMSE of 0.724, MAPE of 10.83%, R2 of 0.918 and pinball loss of 0.247, outperforming all the selected baselines. The 2016 Ecuador earthquake fell within the validation period, so these test results reflect normal operating conditions. These results suggest that condition-sensitive temporal learning can improve demand estimation and provide valuable support for inventory and resilient supply chain planning.