In a recent blog, Intel IT described some lessons learned when developing our Integrated Data Platform (IDP) based on SAP HANA* and Cloudera Distribution of Hadoop* (CDH*). One of the next steps is to show business value by combining this platform, which runs on high-performance Intel® Xeon® processor-based servers, with machine learning.
The overarching principle that guides my supply chain data integration team is transforming reactive scenarios into proactive and predictive scenarios. As long as we are merely reacting—making decisions based on stale data—the business will always be behind. Getting in front of the data is key to creating a “glass pipeline” with end-to-end visibility. And with this glass pipeline and the processing power we have in the platform, we can now predict changes in our supply chain using machine learning.
Intel’s global supply chain spans 100 countries, with over 450 supplier factories and 16,000 suppliers. In addition, Intel fulfills over 1 million orders a year from several factories and 30 warehouses. We do business with many different carriers, and there is a lot of logistics information to process when planning shipments from one location to another.
For example, Intel's finished goods are manufactured both at multiple Intel assembly/test sites as well as by outsourced subcontractors. Finished goods are then forward-staged in several distribution warehouses located in close proximity to key customers. The manufacturing of products is planned based on sophisticated forecasting and planning tools. But forecasting of product movement in the distribution network still involves several manual processes, which is complicated, because Intel ships through multiple distribution warehouses for more than 5,000 SKUs, with fluctuating demand and new product introductions.
This complexity impedes forecast accuracy, and in turn, affects shipment volume and storage space required across the warehouses to handle the uncertainty. The staged warehouses are operated by third-party logistics, whose contract rates are guided by this forecast. So, forecast accuracy directly impacts the cost of logistics operations.
We improved shipment forecasting accuracy by adopting a machine-learning model that uses historical manufacturing output and shipments and manufacturing plans, which are made available through the IDP. The model runs on CDH using Python* and open source libraries to create weekly forecasts at the SKU and warehouse levels for an 18-month horizon; these forecasts are also written back to the IDP to integrate with larger datasets for decision making support.
This graph shows shipment volume predicted by machine learning versus actual shipment volume through a warehouse for a particular SKU. Prior to using machine learning, for a subset of relevant warehouses, forecast accuracy averaged about 70 percent. With machine learning, we are trending to increase forecast accuracy to 80 percent.
Such increased forecast accuracy translates into savings because we do not contract for more shipping capacity than needed. The increased forecast accuracy also helps mitigate the risk of not having the right shipping capacity in place during an upside period, thus minimizing the risk of shipping delays.
This is just one example of the power of near real-time data combined with machine learning. We have many more machine-learning algorithms in development. You can read more about our IDP and lessons learned in the IT@Intel white paper “Transforming Intel’s Supply Chain with Real-Time Analytics.”