Accurately forecasting demand is one of the most undervalued and complex strategies that can significantly impact organizations bottom line. This industrial field study was co-conducted with Sumitomo Corporation's Tubular Division which primarily deals with high-grade Oil Country Tubular Goods (OCTG) globally. The presented solution demonstrates how with the right data set (drilling sequence data, stock data and consumption data), artificial intelligence can be used to build out a model that can quantify and predict future demand accurately thereby reducing cost, working capital and emissions.

Multiple multi-layered machine learning models were built to compare and analyze a wide variety of data inputs for bill of materials, operational/project schedules; This includes (a) ‘product movement data’ which describes the changes in demand and supply of a product, (b) ‘product specification data’ which describes the characteristics of a product, and (c) ‘activity specification data’ which describes the characteristics of an activity. The models follow the base temporal map design with different weighting on model inputs. With a temporal map, a sequence of monthly data values (called lags) is used to predict the next monthly value in the sequence. The lags are rolled so that there are six months of data for the model to predict on. All models also use boosted decision-tree-based ensemble machine learning algorithm.

It is critical to understand how product movement metrics (actual and safety stock levels, historical forecasts, and consumption patterns), product specification data (lead time, product grade, well function, well category, work center), and external factors (oil price, rig counts, national budget, production targets) can be utilized together to better understand future product demand. Using historical data acquired from drilling operations and supply chain over an eight-year period, multiple machine learning models were trained to predict one year of demand across the most consumed products. Across five years of predictions (2016 to 2019), the models were able to predict with 78% average accuracy for the top 10 products by volume which represents 75% of inventory volume. Across the same time-period, they were able to predict with 73% average accuracy on all 17 products which account for 80% percent of inventory volume. Further iterative updates with additional data led to improvement in results and the model where the model predicted with an improved accuracy of 83% on the top 17 products and an accuracy of 86% on the top 10 products.

Moreover, the data can also be used to generate dashboards featuring metrics on material uncertainty / velocity and expected differences between the internally predicted forecasts and actual sales. The results further indicate that, on average, and within a simulated environment (where shipping delays were not considered for instance,) the AI model can maintain a lower inventory than the originally planned stock levels at lowest cost and footprint. This would not only lead to less resource consumption, but also reduce the embodied carbon and emissions within the overall process.

This novel study presents the success of a validated tailored AI model for inventory forecast with field data and commercial implementation. Such a tool can be integrated into other value adding digital tools, such as integrated schedule optimization, logistics optimization and management systems to make overall operations more efficient and sustainable with lower costs, inventory, wastage, and reduced emissions.

You can access this article if you purchase or spend a download.