Anyone who has worked with suppliers recognizes the problem; will the order be delivered on time? Often you do not know until it's too late. In collaboration with LTH and the sister company PipeChain, we started a Master Thesis project that investigates how artificial intelligence can be used to predict the likelihood that the order will arrive on time.
Artificial intelligence (AI) is a hot topic and the possibilities are many. Still, we have many manual decisions and interactions in the supply chains. MA-system together with the sister company PipeChain Group recently completed a Master Thesis with the Industrial Engineering and Management students Emma Ekström and Sofia Danielsson from the Technical Faculty of Lund University (LTH). The aim was to identify common problems in the order process and to develop a practical AI model that could be helpful in the business.
By focusing on order placement in the study, there was a possibility of using data from PipeChain to make a model which predicts when a delivery really comes. The problems are well known by everyone who has worked with purchasing. Different suppliers are reliable to different degrees and delays, as well as premature deliveries, disturbs the business. This leads to, among other things, disruption of planning, production stops and stress. With a model that helps identify and warn if deliveries not will arrive on time, you can work more proactively.
Buildning the model
To build an AI model you need; computing power for calculating combinations and scenarios, an algorithm that selects which data to look at, and data to train the model. By using a model based on a probabilistic approach (Bayesian networks), different probabilities could be calculated for different scenarios. The tool Weka developed at a university in New Zealand was used for the Machine Learning part. There were several types of data available, including:
- The time when the order was placed, it might affect how soon the order is handled or seasonal changes for example
- Time between order placement and desired delivery date, it is difficult to be on time if the goods should be delivered the day after the order is placed
- Quantity, higher quantity may be more difficult to deliver on time
- Price, higher price may be an indication of a complex item.
Over 70% in forecast accuracy
In total, there were over 200,000 data points and 381 suppliers to test on. Different prototypes were developed and tested with varying degrees of success. The final model was tested to work well on two suppliers and then all suppliers as a group. The forecast accuracy was 86.48% for supplier 1 and more than 70% for supplier 2 and the rest of the suppliers. ZeroR is a kind of baseline test to show that the model works, if the model is worse than ZeroR, it's better to use the most common delivery time. The user may see the probability of delivery arriving at different time intervals. A visual interface makes it easy to see which orders that are in the risk zone.
In order for the model to work, it is important to have patterns in the supplier's data. This type of model which is based on historical data, cannot be used with new suppliers since there is no data. The model must first be trained on data to provide good reliability.
Many opportunities for AI in the supply chain
Finally, we can say that there are good opportunities to use AI models to help decision making and to provide better information to a buyer. The combination of human and technology in which the technology provides decision support is in most cases the strongest solution.
From a wider perspective, there are many opportunities to use AI within SCM, som examples are planning, supplier classification, delivery monitoring, and much more. Ultimately, AI will be able to free us from repetitive tasks such as order management and planning to focus on innovation and development.
We wish Emma Ekström and Sofie Danielsson the best of luck in the future and thank them for their efforts with this Master Thesis project.