Improvement of labour planning in industrial manufacturing using Machine Learning

Abstract

Every day, new decisions are made, ranging from trivial to critical, and their impact can vary greatly. Despite these differences, all decisions share a common goal: to choose the best possible action based on the knowledge available at the time. In some cases, past events offer little insight into what lies ahead. However, in many situations, especially in structured environments, historical patterns can reveal valuable information about the future. By analysing these patterns, more informed and accurate decisions can be made.

In production planning, staffing is an important and resource-intensive decision. Assigning the right number of employees to each shift requires balancing cost-efficiency with the need to meet fluctuating demand. This thesis investigates how historical data on production, orders, and shift performance can be used to predict staffing needs. By applying machine learning techniques, we aim to identify patterns that support smarter scheduling and reduce inefficiencies in manufacturing operations. Ultimately, this approach can help companies minimize overtime, optimize resource use, and improve delivery reliability.

To determine the most cost-efficient shift configuration that meets staffing requirements, this thesis first predicted the total number of working hours needed. This was effectively achieved using linear and quantile regression models, particularly when the quantities of individual products were included as separate input features. Modern machine learning models, such as Long-Short-Term Memory (LSTM), also delivered strong results, even when using only aggregated weekly production demand per line. However, given the high performance and simplicity of the linear models, LSTM was considered unnecessarily complex for this specific problem. That said, its ability to capture visible seasonality and reduce autocorrelation suggested that LSTM could still serve as a more flexible option for future use.

After the demand prediction, a cost optimization process was applied to evaluate different combinations of shift-forms and overtime. This enabled the selection of a configuration that not only met the predicted demand but also minimized staffing costs.

Popular Abstract

Modern manufacturing has come a long way since the industrial revolution. Where workers once did everything by hand, machines have today taken over much of the work. This has made it possible to produce more, faster, and more efficiently than ever before. But increased efficiency also brings increased complexity. To truly benefit from advanced machines, production must be carefully planned, and good planning requires good predictions.

Today’s factories are usually organized so that different parts of a product are made at different stations, then assembled into a final product. Each part may require a different amount of time to produce and may be needed in different quantities. Having machines run constantly might seem efficient, but it can quickly lead to overproduction, which costs money in storage. At the same time, it is costly to not fulfil demand as well. That’s why production needs to strike a careful balance: producing just enough of each part while making the most of available workforce hours. This puts significant pressure on staffing decisions, which often require specialized knowledge. Relying on a few key individuals for this expertise can create bottlenecks and risks if that knowledge isn’t broadly shared. Each workstation also has unique requirements, depending on the product mix and the time needed for different tasks. While some durations are easy to estimate, others are more complex, such as machine setup times, recovery after short breaks, or even varying efficiency between day and night shifts.

These patterns are often too complex for traditional planning methods to capture. That’s where machine learning comes in. In this thesis, it has been investigated how both simple and advanced machine learning models can help predict staffing needs in a manufacturing setting. Simpler models like linear regression are useful for identifying straightforward relationships, while more advanced models like Long Short-Term Memory (LSTM) networks can capture both short-term patterns and long-term seasonal trends. The findings showed that both types of models performed well in predicting the demand for working hours, each with their own advantages. Simpler models are easier to understand and use, while more complex ones offer flexibility and potentially higher accuracy, especially with richer data.

To turn predictions into action, an optimization model using Mixed-Integer Linear Programming (MILP) has also been developed. This tool calculates the most cost-effective way to schedule shifts based on predicted demand. It ensures that enough staff are scheduled to meet production needs while minimizing costs, accounting for trade-offs like using overtime when necessary.

In short, by combining machine learning with optimization, we can take a big step toward smarter, more efficient production planning.

• author: Areskoug, Hanna Emma Linnea
• supervisor: Ted Kronvall
• organization: Mathematical Statistics
• alternative title: Optimering av arbetsplanering inom industriell tillverkning med hjälp av maskininlärning
• course: FMSM01 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Long Short-Term Memory (LSTM), Regression, Machine Learning, Mixed-Integer Linear Programming (MILP), Staffing Optimization
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3546-2025
• ISSN: 1404-6342
• other publication id: 2025:E103
• language: English
• id: 9214638
• date added to LUP: 2025-10-30 08:13:39
• date last changed: 2025-10-30 08:16:49