BY MUSTAFA ÇAĞRI GÜRBÜZ
Professor of Supply Chain Management at the MIT-Zaragoza International Logistics Program
A major challenge in supply chain management is matching supply and demand, owing to the uncertainties in the supply process (e.g., random yield, disruptions) and the demand process (e.g., seasonal demand). As integral parts of supply chains, warehouses are critical supply chain actors, serving as a buffer between manufacturers and customers. Even when there is no stock, these facilities are still the primary link between manufacturers and customers, as exemplified by cross-docks. In addition to holding inventory, warehouses aim to reduce costs (inbound/outbound transportation expenses) by consolidating products, achieving economies of scale.
Maintaining reasonably high service levels — shorter response time and high product availability — while keeping the cost-to-serve as low as possible is easier said than done. Reasons for this include shorter product life cycles, the highly perishable nature of products in many sectors (consumer electronics, apparel, fresh produce), SKU proliferation, complex supply chains, and demanding consumers looking for highly customized products. Moreover, as supply and demand processes heavily impact logistics operations, any minor/major disruption to either process would lead to disruptions or considerable inefficiencies in the warehouse.
For example, the increased likelihood of disruptions at the original supplier may incentivize a firm to order from multiple suppliers or decide to keep higher inventories for risk mitigation. These would make inbound processes at the facility more complex and require additional investments to expand storage capacity. Minor disruptions such as delayed deliveries from suppliers and/or short-term demand fluctuations would also make the synchronization of inbound/outbound processes more difficult and cause an unexpected spike in workload.
Predictive analytics can be used to define the Risk Exposure Index of supply chain partners as well as the warehouse itself. It could also ensure that time to survive (i.e., how long a supply chain can endure without a specific node) and time to recover — two well-known concepts introduced by Prof. David Simchi-Levi of MIT — are within allowed ranges to be able to continue business as usual at the warehouse.
Use of data-driven models is increasingly being utilized in supply chain management, especially because decisions can no longer be based on human judgment alone. The availability of large amounts of structured/unstructured data and the improved capability to analyze them allow decision-makers to observe patterns and correlations among different drivers of supply chain performance.
Big data models — descriptive, predictive, and prescriptive — utilize statistical, data mining, and machine learning techniques. Kumar et al. report the use of machine learning models, such as the random forest regression algorithm, used in warehouses, for example, to minimize waste (about one third of global fresh fruits and vegetables become waste and discarded).
Supply chain decision-making is increasingly automated and data-driven
Similar models can also be used for warehouse space/layout planning and location, efficient management of handling equipment and labor, operations monitoring (order picking, inventory, storage), and better decision-making in response to disruptions or unexpected events. DHL, for example, is reported to have used big data analytics technologies for predictive network and capacity planning, customer value management, risk management, real-time local intelligence via pick-up and delivery shipment data analysis, and forecasting of demand and supply requirements. In A machine learning approach for predictive warehouse design, Tufano et al. mention that data-driven algorithms can be used to identify similar SKUs and locate them close to one another or to predict the picking workload and determine zones and picking policies accordingly.
Practical implementations and case studies
WMS with automated identification and data capture
One of the main functions of a warehouse management system (WMS) is to provide visibility between procurement and logistics operations, and the accuracy/quality of data obtained and shared is crucial. Automatic identification and data capture (AIDC) is vital, as it reduces the need for paper-based processes, which lead to reduced input errors.
Digital twins coupled with AIDC and blockchain are promising solutions to enhance efficiency in warehouse operations. Digital twins are not only intended for tracking and monitoring, but also used to build descriptive/predictive models to optimize operations. Internet of Things (IoT) accessibility is an enabler of the development of such tools.
If key performance predictors (KPPs) — e.g., change in customer demand, reduction in inventory levels, change in lead and arrival times of orders from suppliers, handling equipment status, and short-term employee workload — can be predicted with high accuracy, then the WMS together with digital twins would be able to streamline operations. This would eliminate wasted movements (excessive walking distances in the warehouse), increase productivity (workers focusing on other tasks if there are delivery or pick-up delays), and avoid equipment and/or handling material shortage (through predictive maintenance and better manpower planning).
As a result, a WMS optimized with IoT technologies and digital twins would allow businesses to implement strategies such as just-in-time (JIT), vendor-managed inventory (VMI), and cross-docking, which require considerably more effort in synchronization, prediction, coordination, and advanced planning. These systems also have the potential to detect and resolve discrepancies faster. For example, thanks to improved monitoring capabilities and routine information gathering — as opposed to quarterly or yearly audits — the WMS can quickly identify that an item missing from a shipment sent to a retailer was indeed on the warehouse shelf and respond to the claim made by the retailer.
IoT-based WMS to enhance productivity and efficiency
In a study published in the International Journal of Production Research, Lee et al. studied the implementation of an IoT-based WMS integrated with fuzzy clustering techniques to manage the operations of a box build and equipment manufacturing company. Here, customer orders are small and highly customized in a low-volume, high-product-mix environment. This makes traditional manual operations less responsive to changes in orders and even more costly: relying on the memory and experience of workers can amount to 50% of total operating costs for order picking processes.
Customers make frequent change requests, such as pull-in, push-out, and cancellation. And this requires a flexible WMS that would be able to predict such changes and respond promptly to ensure the availability of the raw materials and semi-finished goods involved.
The fuzzy model with a rule-based engine collects information about the current number of orders, number of SKUs, time left until scheduled start date, location of items, customer details, required quantity, and number of workers available to predict the status of the next period. Additionally, it prescribes a response, such as selecting batch-picking as the adequate method over strict order picking.
Therefore, the IoT-based WMS integrated with fuzzy clustering proposed by Lee and coauthors has real-time material monitoring and prompt order-change handling capabilities. Lee and colleagues show that this predictive/prescriptive model improves productivity, picking accuracy, and efficiency and is robust to order variability. The proposed model leads to more orders handled per time unit (less time spent for other fulfillment activities), reduced errors, higher order fill rates, increased order accuracy, and greater inventory record accuracy. These results can be attributed to both the analytical capabilities of the predictive/prescriptive model and the IoT-based WMS.
Predictive warehouse design via machine learning
In the publication Machine learning approach for predictive warehouse design, Tufano et al. developed a machine learning model to predict multiple aspects of a storage system based on previous observations. The first aspect was storage technology, e.g., automated storage and retrieval systems (AS/RS), block stacking, cantilever racks, a mini-load system, pallet racking, and shelving. The second was the material handling system (cart, forklift, operator, order picker, etc.). The third involved the storage allocation strategy, for instance, reserve & forward policy or simple storage without duplication (i.e., reserve policy). The fourth aspect was the picking policy, for example, single- or multi-order.
Big data technologies and machine learning allow businesses to implement descriptive, predictive, and descriptive models
The KPIs of the model are SKU profiling (classifying the behavior of each SKU), inventory profiling (understanding the behavior of the saturation of the space), workload profiling (identifying where and how the workload is distributed), and layout profiling (determining how resources are placed/organized). Inventory profiling, for example, would be used to predict the risk of stockout, identifying the time before the inventory is consumed by the market demand for a particular SKU.
This model requires various inputs including inbound data (putaway), outbound data (picking), layout information, layout coordinates, volume data for each SKU, and picking list details. The model aims to predict the warehouse configuration to assign to each SKU and has been validated with data collected from 16 companies in industries such as automotive, manufacturing, food and beverage, cosmetics, and publishing.
The model has distinct practical implications for different supply chain players and aims to provide feasible — not necessarily optimal — solutions with high flexibility given the industrial practices. For example, 3PL providers generally find it more difficult to make such design choices because the demand they face is more unpredictable, often characterized by lumpy patterns. Moreover, they may not know the exact needs of new customers, and existing contracts may change frequently. An example would be the rotation of SKUs owing to the expiration of client contracts. Tufano and colleagues report that these 3PLs can potentially benefit from this data-driven approach, especially if information on crucial parameters such as the volume/weight of each SKU and the dynamics of the market demand (popularity, seasonality) is made available.
Short-term workload forecasting and effective labor management
Supply chain decisions have been increasingly data-driven and automated. However, human judgment is still a critical factor in supply/demand planning. Furthermore, a significant portion of warehouse operations, namely order picking and packing, are still labor-intensive. Warehouse workloads — especially outbound operations — are also quite variable, mostly because of the uncertainties in supply and demand processes.
In navigating demand fluctuations, businesses must adapt to several factors, including seasonal demand for certain product types and/or peaks in end-of-period demand due to incentives such as sales targets. On account of these challenges and the resulting variability in the workload, companies look for flexible labor pools in addition to permanent full-time personnel in the warehouse. Consequently, forecasting workload and capacity planning (i.e., warehouse manpower) requires a careful analysis aimed at detecting human judgment in demand planning, controlling any potential bias, and estimating the impact on labor efficiency.
Controlled optimal forecast bias relates to human judgment leading to over-forecasting of order sizes by managers. This bias arises from the cost implications of over/under forecasting, which depend on labor hiring options and service level agreements with customers. While this bias does not seem to benefit the labor-intensive packing stage, a study at a Samsung Electronics warehouse in Western Europe for fast-moving goods found that a bias of around 30-70% in picking and loading operations resulted in order efficiency gains of 5-10%. The findings were also validated by a survey of 30 warehouses belonging to other companies (see Kim et al.).
The predictive model proposed by Kim et al. confirms that controlling bias can improve warehouse labor efficiency. The model defines the analytic relationship between demand forecast bias (the difference between forecasted demand and actual order size) and labor productivity. This correlation is then leveraged to optimize labor capacity planning.
Integrating expert judgments with historical demand data and controlling the forecast bias by correcting recently observed biases improve sales forecasts and allocation of labor capacity across different stages.
Predictive models improve warehouse management
Big data platforms that would allow companies to deploy descriptive, predictive, and prescriptive models can be used to mitigate various risks, increase efficiency, and maximize profitability in warehouse management, so long as the barriers are overcome. Organizational size limited technological resource competency, and lack of information sharing with supply chain partners also pose challenges to big data analytics adoption in warehouse management.
Ghaouta and colleagues’ research suggests that the use of predictive models remains limited — most models being descriptive — except for routing algorithms and inventory control. However, increased awareness in supply chain risk management will most likely encourage businesses to adopt predictive models more often in warehouse management. A WMS enhanced with risk management capabilities could identify and assess the potential risks (likelihood of disruptive events occurring and the potential impact) associated with supply chain partners, such as suppliers prone to disruption or customers with erratic ordering behavior. It would further develop reactive and/or proactive mitigation strategies based on segments of suppliers/customers with different risk profiles.
More research on how the IoT-based WMS would allow companies to make the transition from centralized to decentralized control is also necessary. Humans, handling materials, products, and sensors would be well connected in such systems. And this has the potential to make the improved coordination and communication of decisions a possibility.
- Kumar, M.N. Vimal, S. Snehalatha, C. Shobana Nageswari, C. Raveena, and S. Rajan. 2021. Optimized Warehouse Management of Perishable Goods. Alinteri Journal of Agriculture Sciences 36 (1): 199–203.
- Ghaouta, Ayoub, Abdelali El bouchti, and Chafik Okar. 2018. Big Data Analytics Adoption in Warehouse Management: A Systematic Review. 2018 IEEE International Conference on Technology Management, Operations and Decisions (ICTMOD), November.
- Tufano, Alessandro, Riccardo Accorsi, and Riccardo Manzini. 2021. A Machine Learning Approach for Predictive Warehouse Design. The International Journal of Advanced Manufacturing Technology 119 (3-4): 2369–92.
- Smith, Alan D. Warehouse Management Systems: Comparison of Two Pittsburgh-Based Manufacturing Firms. In Encyclopedia of Information Science and Technology, Sixth Edition, edited by Mehdi Khosrow-Pour, D.B.A. Published ahead of print, 3/22/2023.
- Lee, C.K.M., Yaqiong Lv, K.K.H. Ng, William Ho, and K.L. Choy. 2017. Design and Application of Internet of Things-Based Warehouse Management System for Smart Logistics. International Journal of Production Research 56 (8): 2753–68.
- Kim, Thai Young, Rommert Dekker, and Christiaan Heij. 2018. Improving Warehouse Labour Efficiency by Intentional Forecast Bias. International Journal of Physical Distribution & Logistics Management 48 (1): 93–110.
Dr. Mustafa Çağri Gürbüz is a Professor of Supply Chain Management at the MIT-Zaragoza International Logistics Program. He is also a Research Affiliate at the MIT Center for Transportation and Logistics. His main research interests are inventory and supply chain management, distribution system optimization, contracts, and operations systems modeling.