Advancing Dynamic Flexible Job Shop Scheduling With a Real-World Dataset and Genetic Adaptive Scheduling System

In this study, we built a large-scale real-world dataset for further exploration of the Dynamic Flexible Job Shop Scheduling Problem (FJSSP) in a complex production environment that was developed and conducted based on a Genetic Adaptive Scheduling System (GASS). In contrast to traditional static datasets, the dataset we present captures a variety of operational and technical variables from the industrial location of actual flexible packaging manufacturing, such as machine capacities, job priorities, setup time, processing speed, routing flexibility, and unplanned machine breakdowns. The dataset is based on logs extracted from enterprise resource planning (ERP) systems and reviewed by industry experts, providing a realistic ground truth for simulating and evaluating the performance of our dynamic scheduling strategies. This dataset acts as the input to the GASS model, which generates near-optimal and adaptive schedules using genetic algorithms with informed initialization, multi-objective fitness evaluation, and rolling horizons. Extensive Experimental results showed significant performance improvements in terms of total penalty, makespan, and workload imbalance compared with the traditional genetic algorithm and dispatching rules, with reductions of 18.7%, 24.3%, and 21.5%, respectively. This work is distinct from our prior publications in that it released an industry-validated open dataset for dynamic FJSSP with full schema and FAIR metadata, together with a permissive license and loaders for immediate reuse. In addition to prior GA-based studies, we introduce an adaptive GASS mechanism—guided initialization and rolling-horizon re-optimization triggered by urgent job arrivals and machine downtimes, and validate it on real factory data rather than simulations.