Probabilistic Demand Forecasting for Intelligent Industrial Convergence: An Enhanced Informer-GRQLSTM Framework for Multi-Category Supply Chains
Main article
Abstract
Demand forecasting in complex supply chains has moved from a routine planning exercise to a core capability for resilient industrial decision-making. The uploaded source manuscript proposes a hybrid architecture that combines an Informer encoder with a gated residual quantile long short-term memory decoder to improve multi-category forecasting under uncertainty. Building on that manuscript, this rewritten article reorganizes the study into a clearer systems-oriented research narrative and expands the analytical discussion for the Journal of Intelligent Industrial Convergence format. The proposed framework integrates temporal representation learning, residual gating, quantile regression, and sequence memory to model nonlinear demand, seasonal fluctuations, and probabilistic forecast intervals. Using a real-time supply chain dataset covering 180,520 shipments from 2018 to 2023 across clothing, sports, and electronics categories, the study evaluates the model through deterministic and probabilistic metrics. Reported results indicate strong forecasting performance, with MAE of 0.0165, MSE of 0.0178, RMSE of 0.0121, SMAPE of 0.0172, Q-Risk of 95%, and Winkler Score of 0.127. Compared with TS-SimPMF, TCN, attLSTM, CEEMD, UWDFNET, and SARIMA baselines, the hybrid framework achieves lower error, shorter inference time, and better stability under ablation testing. The paper further extends the original discussion by examining industrial applicability, regional demand heterogeneity, robustness, and managerial implications for replenishment, capacity control, and inventory buffering. The results suggest that probabilistic deep forecasting can support intelligent industrial convergence not only by improving predictive accuracy, but also by aligning data-driven anticipation with operational coordination across production, logistics, and market response.
