Neural networks for forecasting raw material demand
At Tasowheel, collaboration with local universities is a strategic tool for generating new knowledge and staying in the front row of development. Student projects and theses play an important role in collaboration.
Recently, computer science student Santeri Kuusisto’s master’s thesis about forecasting raw material demand in the manufacturing industry was published at Tampere University. As part of his research project, Santeri designed a neural network-based forecast tool for predicting the total demand for raw materials. His passion for machine learning and forecasting, together with long co-operation with Tasowheel, inspired Santeri to start creating a practical solution that would bring real value to the company.
For production manager Mikael Mäkinen, the number one objective was to have a forecasting tool that would help Tasowheel give their suppliers more accurate and on-time estimates of upcoming material needs. Machine learning seemed to open up new possibilities of which he was eager to learn more. Good experiences with Santeri developing Tasowheel’s reporting convinced him that he was the right person to do the job.
Let’s find out how Santeri succeeded!
How does neural network-based forecasting work?
Neural networks are a set of algorithms that are designed to recognize patterns. We upload data input, such as the historical data of the raw material demand, the number of products using the raw material in question, and the company’s turnover. All data is directly gathered from the enterprise resource planning system, and the final forecasts are visualized with Power BI.
What is the key take-away from your research?
Neural networks tend to produce much more accurate forecasts than estimating techniques that are based on the previous period’s actuals, i.e. naive forecasts. In other words, the prognoses turned out very good!
How can a manufacturing company benefit from your forecast tool?
The end-users are purchasing managers responsible for procuring raw materials. This will give them a tool to forecast the optimum level of raw materials inventory and, ideally, lead to more efficient inventory management.
What makes raw material forecasting tricky?
There is a great monthly variation in raw material demand, so even a good forecast may have an average error of more than ten percent.
Could the forecast tool be developed further?
The tools that we used are suitable for creating prognoses within neural networks, but their calculation time was rather long. If there is more data to be analyzed it might be recommended to use cloud-based calculation services for making shorter calculation times.
Now that the forecast tool is in use, what will be your next steps?
First, it will be interesting to see if the forecasts bring more efficiency to inventory management. The next step will be to investigate how richer input data, such as an index of general economic conditions, affect forecast accuracy.