Misperformance Forecasting for Logistics

Project description

The goal of the project is to create an AI-powered system that can be used by KLG's shipment planners to forecast misperformance for the shipments they organize. In other words, we aim to predict if a shipment will arrive later than scheduled, and how many days the delay will be.

Context

KLG Europe is a Dutch logistics and transportation company that specializes in a wide range of supply chain services, primarily focused on freight forwarding, warehousing, and distribution. Founded in 1918 in the Netherlands, KLG Europe has grown into an international player with a strong presence in Europe and beyond, offering end-to-end logistics solutions. KLG Europe offers a variety of transport solutions, most notably by road (trucking), which is most relevant for the scope of this project.

KLG's planners are integral for the company's operation, and they are important stakeholders for us. On daily basis, they work around the clock to organize deliveries for their numerous customers in Europe and beyond. KLG aims high for their KPIs - up to 98% of the shipments need to be delivered on time. This is a high target, considering the unpredictable nature of the domain. Constantly, there are circumstances which are outside of KLG's control, such as traffic jams, extreme weather events, protests, accidents, trucks malfunctioning, etc.

In our interviews with the planners, we learned that all of them utilize one precious tool above anything else - their intuition, forged by years of experience in the field. While it is remarkable how they can be so efficient by following their gut, there are certainly downsides to this - intuition takes years to develop, and the know-how of a planner is lost when he is absent. For that reason, KLG wishes to make planners' life easier by utilizing in-house AI which predicts when deliveries will be late.

Results

In the very beginning of the project we had to familiarize ourselves with KLG's domain. We were provided some data to explore, and we held several meetings with the stakeholders, who made sure to give us a proper introduction into the world of logistics. After having some foundational understanding of the specifics, we started getting deep into the data. We analyzed correlation and causality, created interactive visualizations to help us understand how the goods move across Europe, and enriched our data with third-party datasets which allowed us to engineer valuable data features to be later used by our AI models.

The data analysis we produced was found very insightful by the stakeholders, as it allowed them to view their operation from a different perspective, which had not been explored before.

At that point, we started prototyping our first models - we organized our workload into stand-alone tasks, allowing the four of us to branch out and simultaneously explore different approaches. We tried out deep learning as well as classical ML algorithms, and after a fair amount of experimentation, feature engineering and validation, we had several promising candidate models which we could focus on improving. We ended up building two separate neural networks, using the same data, although processed in slightly different ways. One of the them was a classifier, simply predicting whether or not a delivery will be late, and the other was a regression network, which outputted the number of days of expected delay for the delivery.

We were able to produce predictions with close to 90% accuracy for the classifier model and sub 1 day error for the regression model. Both models were made accessible via a web interface, which will allow KLG's planners to easily make predictions for the deliveries on their agendas.

About the project group

We're a group of four ambitious students, coming from a software engineering background, working together to build a system to help KLG's shipment planners do their job more efficiently. The project spans an entire semester and we usually spend 2 to 3 days per week working on it.