Introduction
This study focuses on improving the efficiency of automated warehouse order picking by addressing the challenge of predicting how long it takes to complete a group of picking orders, known as the makespan. The system under analysis is a multi-level shuttle warehouse, a storage and retrieval system that handles goods without human intervention.
The authors propose a neural network-based metamodel designed to predict the makespan. This metamodel is trained using data from a discrete-event simulation model developed by AnyLogic.
Simulation model
The authors developed a discrete-event simulation using AnyLogic to train the predictive model. This simulation represents a realistic multi-level warehouse order picking system, including features such as:
- Multiple aisles and storage locations.
- Handling machines that move goods along horizontal and vertical tracks.
- Picking stations where items are collected.
- Conveyor systems that move storage units.
The simulation captures the flow of orders, movements of machines, and handling of product units, including minor machine breakdowns. It does not consider incoming stock (replenishment) or shipping processes, focusing only on picking.
Key features of the AnyLogic model:
- Simulated 500 different products stored in various quantities.
- Considered machine delays and mechanical errors.
- Tracked how long it took to complete all items in customer orders.
Results
In total, the authors generated 3,500 unique sets of customer orders. Each was run 100 times to account for randomness, resulting in a dataset of 350,000 samples used to train the machine learning model.
Also, read our blog post to learn how machine learning in simulation transforms planning, testing, and optimization of complex processes.
The trained neural network model was evaluated on new data and showed strong accuracy in predicting the total order completion time:
- The average prediction error was about 176 seconds.
- The percentage error was around 5.4%, indicating reliable performance.
The simulation built in AnyLogic successfully captured the complex behavior of the warehouse order picking system and provided a strong foundation for training a predictive model. The result is a practical tool for warehouse managers to make quick decisions.