An Integrated Deep Learning and Multi-Objective Pareto Optimization Framework for Retail Supply Chains
Ziyad Ahmad Mohammed, Anas Chafi, Mohammed El Hammoumi · IEEE Access · 2026
Modern supply chain environments are characterized by stochastic demand volatility and multifaceted operational interdependencies that render traditional single-objective optimization models inadequate. This study proposes an integrated methodological framework that synthesizes advanced deep learning architectures with multi-objective Pareto optimization to reconcile conflicting goals of cost-efficiency and service reliability. We implement a hierarchical ensemble comprising Quantile Regression, Bidirectional Long Short-Term Memory (Bi-LSTM) networks, and Transformer models equipped with attention mechanisms to capture complex non-linear temporal dependencies across a large-scale retail dataset of 913,000 observations. The ensemble weights are determined through cross-validation-based optimization on a held-out validation set. The proposed weighted ensemble achieves a superior $R^{2}$ of 0.9299, representing a 23.8% improvement over the naive baseline, while also outperforming state-of-the-art models including N-BEATS ( $R^{2}=0.9185$ ), TiDE ( $R^{2}=0.9210$ ), and PatchTST ( $R^{2}=0.9245$ ). These high-fidelity forecasts are subsequently integrated into a multi-objective decision-support engine that concurrently optimizes four explicitly defined objectives: inventory holding costs, service levels, delivery efficiency (measured as reduction in store-level order variability), and production smoothness (measured as reduction in daily production variability). Each objective is formulated with complete mathematical definitions and decision variables. By identifying non-dominated Pareto-optimal solutions, the framework allows decision-makers to navigate the strategic trade-off space. Empirical validation demonstrates delivery efficiency improvement of 70.7% and production smoothness improvement of 70.2%, defined as normalized percentage reductions in standard deviation relative to the naive forecasting baseline. Furthermore, financial analysis projects conservative annualized savings of ${\$}$ 5.26M, reflecting a Return on Investment (ROI) of 1,051.9% and a payback period of 1.1 months. The implementation code is publicly available on GitHub for reproducibility.