Digital Twin For Warehouse Resource Planning

Problem Statement

In today’s digital age, e-commerce companies are constantly seeking ways to gain an edge over their competitors. With customers expecting faster and more convenient services, companies are offering increasingly attractive options. One such offering is one-day or same-day delivery. While these services are certainly attractive to customers, they put enormous strains on their delivery partners.

Many small e-commerce players rely on third-party aggregators to handle their logistics operations. These companies establish a Service Level Agreement (SLA) that defines certain Key Performance Indicators (KPIs). When these KPIs are unmet, penalty is levied by merchants. As a result, companies make every effort to avoid breaching their terms of SLAs.

In a warehouse, one such KPIs is total order cycle time, i.e., the average time it takes for an order to get shipped, starting from the moment when the order is received. Once an order is placed at a merchant site, a corresponding request is generated in a warehouse where it goes through multiple stages. The order needs to be located and picked from its rack, which is usually done in batches to save time. However, this raises multiple “what if” scenarios. For example, what if some customers have overlapping products? What if only one order was placed? What if we have multiple items that are of different categories and require different packaging? These complexities are generally handled by the warehouse manager, by maintaining Standard Operating Procedures. After the order is picked, it needs to be sorted, packaged, and prepared for courier pickup. These processes require human effort, and the manager needs to manage resources at each stage based on the order flow.

Although future order flow can be estimated by analyzing historical data and resources can be planned based on these forecast, but there can still be unexpected events that can compromise the entire process. So it is of utmost importance to have tools that can help manage warehouse resource planning in advance. The solution should be able to simulate all possible scenarios and generate resource allocation plan. It should also be able to keep up with the physical changes and changes in SOPs, for example if an improvement is made in the warehouse to reduce pick up time, it should reflect in the system.

Solution

Due to the presence of multiple random factors that are connected to various uncertain events a deterministic solution would not suffice. These uncertainties introduce a level of complexity that makes it challenging to predict whether the Service Level Agreement (SLA) will be achieved or breached. To tackle this, a modeling approach known as Monte Carlo simulation can be utilized.

Monte Carlo simulation is a technique that involves running numerous iterations of a model by randomly sampling input values from the probability distributions. In the context of our digital twin, Monte Carlo simulation can help generate a range of possible outcomes for meeting SLAs by considering the variability and uncertainty in factors such as order flow, time taken by workers to complete each task, variability in operational hours, and other relevant variables. Our ultimate goal is to achieve SLAs even in worst case scenarios.

Other application of the tool is to simulate What if? scenarios. What if we have X number of resources and Y number of order, how should we distribute these X at each assembly to minimize order cycle time? What if we have very large Y orders, how many resources we need to acheive SLAs? Various permutations and combinations can be input to the model and based on these inputs other dependent variables can be calculated to maximize or minimize any specific metrics.

Methodology

Just like in any problem within the data science field, the initial step involves curating and vetting the data in our system. Ensuring the quality and reliability of the data is crucial for accurate analysis and modeling. After the data is vetted, we explored it from various angles. Specifically, for analyzing the historical order pattern, we examined factors such as the frequency of orders based on the day of the week, week of the month, holiday season, and product type. This analysis guided our selection of features for modeling the probability density distribution.

While opting for a Gaussian Distribution may simplify the process, it is not always the most suitable estimator. In our case, we found that a Gaussian Distribution was not a good fit for our data. Consequently, we employed Kernel Density Estimate, a non-parametric method for estimating the probability distribution function. This approach allowed us to capture the underlying probability distribution accurately.

A similar approach could be used to estimate other random factors, like the time taken by resources to pick up specific types of orders, sort multi-batch orders, pack different kinds of items, and so on. This was again analyzed on multiple longtitudinal aspects. During our analysis, we discovered that even the day of the week has an impact on resource performances. This finding highlights the importance of considering various temporal aspects when examining resource efficiency and identifying patterns that can influence operational performance.

Once all the random factors were modeled, we leveraged Monte Carlo simulation to draw independent samples from the respective probability distributions. This simulation allowed us to replicate and analyze all stages of the warehouses in a controlled environment.

The iterative nature of the simulation enabled us to generate comprehensive reports for worst-case, average-case, and best-case scenarios. These reports included valuable insights such as resource requirements, productivity, on-time shipping, average order cycle time and other key performance indicators (KPIs).

Result

By examining the results from the simulations, the warehouse managers can gain a holistic view of potential outcomes and be able to plan their activities accordingly. This information can be proved instrumental in optimizing resource allocation, enhancing operational efficiency, and making informed decisions to meet the desired objectives.