A Dynamic Adaptive Fusion Transformer for Zero-Inflated Intermittent Spare Parts Demand Forecasting

In the complex world of supply chains, keeping the right parts in stock is a constant balancing act. If a factory or an airline runs out of a specific component, operations can grind to a halt, causing costly delays. But if a warehouse is filled with parts that never get used, money sits idle on shelves, and the items might eventually become obsolete. This challenge is especially difficult for spare parts used in machinery and vehicles. Unlike the demand for everyday items like bread or batteries, which people buy regularly, spare parts often sit untouched for months or even years, only to be needed suddenly when a machine breaks. This pattern, where long periods of silence are interrupted by rare, unpredictable requests, creates a unique forecasting problem. Standard prediction tools, which work well for steady sales, often fail here. They tend to guess that a part will be needed soon simply because the math says so, leading to a buildup of unnecessary inventory, or they miss the rare moment when a part is actually required.

A team of researchers from the Air Force Logistics College in China has developed a new approach to solve this specific puzzle. They created a sophisticated computer model designed to handle these erratic, "zero-inflated" demand patterns, where most months show no demand at all. The researchers tested their system using a real-world dataset containing demand records for over 2,600 different spare parts, tracking their usage over a period of more than four years. The goal was to predict whether a part would be needed in the coming months and, if so, how many units would be required. Their new model, which they call the Dynamic Adaptive Fusion Transformer, works by splitting the prediction task into two distinct questions: first, will the part be needed at all? and second, if it is needed, how many will be required? By treating these as separate problems and then carefully combining the answers, the model learns to recognize the difference between a true lack of need and a rare spike in demand.

The researchers found that this split approach, combined with a mechanism that allows the model to adjust its own confidence, significantly improved accuracy. In their tests, the new model was far better at avoiding false alarms than previous methods. When the actual demand for a part was zero, the new model predicted a near-zero need, effectively stopping the warehouse from ordering parts that would just sit there gathering dust. Specifically, the model reduced the error in predicting zero-demand months to a very low level, far outperforming traditional statistical methods and other modern machine learning tools. However, the study also revealed a trade-off. While the model became excellent at knowing when not to order, it occasionally became slightly more cautious when predicting large, sudden spikes in demand. It successfully suppressed the noise of false positives but sometimes underestimated the size of a genuine, urgent request.

To understand how this works, imagine the model as a system that first checks a calendar and a history log to decide if a request is likely. If the history shows the part hasn't been used in a long time, the system leans heavily toward predicting zero. If the history shows recent activity, it shifts its attention to estimating the quantity. The researchers built a "gating" system that acts like a traffic controller, deciding how much weight to give to the "will it happen?" guess versus the "how many?" guess. This allows the model to be flexible. In months where the data suggests a quiet period, the model confidently predicts zero. In months where the data hints at activity, it opens the door to predicting a positive number. This dynamic adjustment is what allows the system to handle the extreme unpredictability of spare parts without getting confused by the long stretches of silence.

The experiments showed that this method is particularly effective for parts that are rarely used but critical when needed. By reducing the number of times the model incorrectly predicts a need for a part that isn't actually needed, the system helps organizations save money on storage and avoid the waste of ordering obsolete stock. The researchers noted that while the model is not perfect at predicting the exact size of a massive, sudden demand spike, its ability to avoid false alarms is a major improvement. In the context of inventory management, where the cost of holding extra stock can be high, this precision is valuable. The study suggests that for many spare parts, the best strategy is to be very careful about predicting a need unless the evidence is strong, rather than guessing that demand might appear just to be safe.

The researchers tested their model against several other methods, including older statistical techniques and newer deep learning models. The results were clear: the new model produced fewer total errors and was much better at identifying months where no demand would occur. In fact, when looking specifically at the months where no parts were needed, the new model's predictions were significantly closer to zero than those of the other methods. This means it successfully avoided the common pitfall of "spurious positive predictions," where a model sees a pattern that isn't there and orders parts unnecessarily. The study also confirmed that each part of their new system played a vital role. Removing the part that checks for demand occurrence made the model worse at handling zero-demand months, while removing the part that checks for demand size made it less accurate when parts were actually needed. The combination of these elements, managed by the adaptive fusion mechanism, created a system that was more robust than any single component alone.

While the model excels at preventing over-ordering, the researchers acknowledged that it still has room to improve when it comes to predicting the magnitude of rare, high-volume events. In their tests, the model sometimes predicted a smaller number of parts than were actually needed during a sudden surge. This suggests that while the system is great at knowing when to stay quiet, it can still be a bit conservative when the noise of a real demand spike arrives. The authors propose that future work could involve adding specific rules to handle these extreme peaks or combining the model's output with safety stock policies to ensure that critical parts are never missed. For now, the study demonstrates that by breaking down the forecasting problem into smaller, more manageable pieces and letting the model decide how to weigh them, it is possible to create a much smarter system for managing the unpredictable world of spare parts.

The implications of this work extend beyond just the military or aviation sectors, where the researchers are based. Any industry that relies on maintaining equipment with parts that are used infrequently could benefit from this approach. From industrial manufacturing to automotive repair, the challenge of predicting intermittent demand is universal. The success of the Dynamic Adaptive Fusion Transformer suggests that the future of inventory management lies not in trying to force a single, simple prediction onto complex data, but in building systems that can understand the different states of demand and adapt their strategy accordingly. By learning to distinguish between a true lack of need and a rare moment of high activity, these systems can help businesses operate more efficiently, keeping their supply chains lean without risking the availability of the parts they need most.