A Data-Driven Methodology for Hierarchical Production Planning with LSTM-Q Network-Based Demand Forecast

Abstract

Mass customization makes it necessary to upgrade production planning systems to improve the flexibility and resilience of production planning in response to volatile demand. The ongoing development of digital twin technologies supports the upgrade of the production planning system. In this paper, we propose a data-driven methodology for Hierarchical Production Planning (HPP) that addresses the upgrade requests in the production management system of a fuel tank manufacturing workshop. The proposed methodology first introduces a novel hybrid neural network framework with symmetry that integrates a Long Short-Term Memory network and a Q-network (denoted as LSTM-Q network) for real-time iterative demand forecast. The symmetric framework balances the forward and backward flow of information, ensuring continuous extraction of historical order sequence information. Then, we develop two relax-and-fix (R&F) algorithms to solve the mathematical model for medium- and long-term planning. Finally, we use simulation and dispatching rules to realize real-time dynamic adjustment for short-term planning. The case study and numerical experiments demonstrate that the proposed methodology effectively achieves systematic optimization of production planning.