Contextual distributionally robust optimisation for inventory replenishment optimisation under uncertain demand
提出结合Transformer神经网络与共形推断的情境分布鲁棒优化框架，利用电力物资数据验证，相比传统方法效率提升56%-69%，帮助管理者应对需求不确定性。
Demand uncertainty poses a significant challenge in inventory management, where conventional strategies often fail to adapt to complex and fluctuating demand patterns. To address this critical issue, we propose a novel contextual distributionally robust optimisation (CDRO) framework that integrates advanced machine learning with robust decision theory. Our solution features the HOA-CITransformer predictor, which combines a Transformer neural network optimised via the Hiking Optimisation Algorithm (HOA) with Conformal Inference (CI) to generate accurate demand forecasts (95.6% accuracy) with reliable uncertainty quantification (88.9% coverage probability). Building on these probabilistic predictions, we develop a Wasserstein distance-based CDRO model that constructs ambiguity sets to ensure robust inventory decisions under demand variability. Experimental validation using power grid material management data (2020-2023) demonstrates the framework's effectiveness: it achieves efficiency improvement by 56.15%−69.29% with high guarantee of satisfaction rate compared to traditional methods. This research provides practitioners with a principled approach to overcome demand uncertainty challenges in supply chain operations.