Adaptive Deep Learning Ensemble for Supply Chain Demand Forecasting

Accurate supply chain demand forecasting is critical for inventory optimization and risk reduction. This study proposes three enhanced forecasting models, specifically Dynamic Weight Fusion-based Uni-Regression Deep Approximate Forecasting (DWFUDAF), Multi-Head Attention-based UDAF (MHAUDAF), and a combined Dynamic Weight Fusion-MHA model (DWFMHAUDAF). These models are built upon Bidirectional Long Short-Term Memory (BiLSTM) and Nonlinear Autoregressive models with exogenous inputs (NARX) to address the limitations of fixed-weight neural-ensemble architectures such as UDAF. By introducing Dynamic Weight Fusion (DWF) and multi-head attention (MHA), the proposed architectures adaptively reflect temporal demand shifts. The proposed models are evaluated through a focused methodological validation using a highly volatile time-series (Store 1) from the Rossmann Store Sales dataset. This specific store was selected as a complex testbed to rigorously assess the adaptive capability of the DWF mechanism under dynamic demand shifts and exogenous influences. Empirical validation on the Rossmann Store Sales dataset, benchmarking ten forecasting models, showed that DWFUDAF achieved the lowest Mean Absolute Error (MAE = 0. 147). In contrast, the statistical ARIMAX baseline recorded the lowest Root Mean Squared Error (RMSE = 0. 206). Statistical tests, specifically one-way ANOVA and Tukey's HSD, confirmed that DWFUDAF outperformed both fixed-weight and attention-based ensemble architectures in MAE (p < 0. 001). In contrast, MHA-based models exhibited degraded accuracy. This study attributes this performance drop to the structural mismatch between the high data requirements of Multi-Head Attention and the limited sample size of a single store, which led to overfitting and attention weight collapse. This work provides both theoretical and practical evidence that adaptive-fusion mechanisms outperform static structures in complex supply chain forecasting scenarios.