Australia Post - Leveraging ML-based Forecasting To Optimize Capacity Planning at Processing Facilities in a Large-scale Logistics Network
James Walter, Senior Data Scientist Yohan Ko, Senior Data Engineer Btou Zhang, Network Operations Lead Duc Nguyen, Shift Production Manager Normy Chamoun, Head of Processing NSW/ACT Sheral Rifat, Data Science Manager Phil Chan, Data Engineering Manager David Barrett, Facility Manager Boris Savkovic, Data Science Manager
Organization: Australia Post
Australia Post is a government business enterprise that provides postal services in Australia. We are also Australia’s leading logistics and integrated services business. Last year, we processed 2.6 billion items, delivered to 12.4 million delivery points across the nation, and continued to provide essential government and financial services via the country’s largest retail network.
Value at Scale
Most Impactful Transformation Story
Most Extraordinary AI Makers
We are Australia’s leading logistics and integrated services business. Last year, we processed 2.6 billion items, delivered to 12.4 million delivery points across the nation, and continued to provide essential government and financial services via the country’s largest retail network.
The global pandemic has accelerated e-commerce growth with more households shopping online than ever before. Whilst Australia Post has a long and proud history to lean on, we continue to face challenges from ever-increasing parcel volumes and a great digital disruption that is shaking up the wider logistics industry. This requires us to innovate and transform.
A key daily activity, at facilities within our logistics network, relates to shift production managers being tasked with making daily resource/staffing planning decisions, that seek to ensure that we process parcel demand in a timely manner, whilst controlling for cost. Currently, these decisions are being made based on limited, but best-available, information. Too few staffing hours can result in sub-optimal throughput and parcel delays, whilst too many staffing hours can unnecessarily increase labor spend.
To address this pain paint, our Data Science team developed a shift volume forecasting algorithm in Dataiku. The model provides facility operators with daily shift volume forecasts and translates this information into staffing requirements.
The algorithm was trialed in partnership with one of the biggest processing facilities in the Southern Hemisphere, Sydney Parcel Facility, and is now used to inform daily planning activities.
Feedback from shift production managers is that "based on the volume prediction, we were comfortable with not running overtime the following morning. This paid off". Thus the model is empowering managers to confidently make decisions regarding the need for overtime.
The approach is changing the way that facility operators make decisions, resulting in significant operational dollar value savings (~15 million Australian Dollars [AUD] p.a. once rolled out nationally).
We chose Dataiku as we were looking for an end-to-end data science platform that simplified and automated many aspects of the data science and data engineering workflow, allowing our team to deliver results faster and with fewer frustrations.
The team made use of Dataiku starting with initial exploratory data analysis (EDA), for python coding and use of custom modules, through to production deployment and BAU operation including model performance monitoring, data monitoring, and resource monitoring.
The end-to-end MLOps process in Dataiku is streamlined, integrated, and easy to use. Specifically, Dataiku has enabled us to easily manage the following key aspects of the MLOps workflow:
Complex dependencies in terms of libraries and virtual environments, by abstracting many of the complexities that one usually faces when working with dockerization or virtual environments.
Scalability of our models by providing a streamlined way to leverage a Kubernetes cluster on GCP to attain scale and to enable further scale-out of the model to future facilities.
Version control functionality in Dataiku Data Science Studio (DSS) enables lineage tracking and model versioning across the full lifecycle.
Collaboration functionality whereby data scientists, data engineers, and developers could co- develop and then serve the models seamlessly to business users.
ETL pipeline development, leveraging both time-based and event-driven scenario execution, to process the real-time data that is feeding into our models.
We rapidly developed an ML model (random forest model in Dataiku, leveraging the Visual ML capability) to forecast shift volumes and labor/staffing requirements at each facility.
Model deployment to production in Dataiku was speedy and required minimal resources from data engineering as many of the production processes are automated and handled by the Dataiku platform.
The metrics, checks, and testing capabilities have enabled us to add quality assurance to our models
Business Area: Supply-chain/Supplier Management/Service Delivery
Use Case Stage: In Production
This project is highly innovative, novel, and transformative within Australia Post, as it is bringing real-time forecasts to users at our facilities, thus enabling a level of real-time and data-driven decision-making that has not been possible to date.
In short, operational decisions can now be made in a timely manner as required by the time-constrained daily cycle of our network operation teams.
Most importantly, these forecasts are relevant, actionable, accurate, and highly automated through the use of the Dataiku platform.
From a scale and technical point of view, the forecast generation process is now streamlined end-to-end in Dataiku, and can easily be scaled out to more facilities nationally.
Specific business metrics of success include:
The data-driven forecasting approach is changing the way that facility operators make decisions, resulting in significant operational dollar value savings (~15 million AUD p.a. once rolled out nationally), and significant uplifts in overall parcel throughput within our network. The dollar value savings result from reduced labor costs at facilities (reduced spending on on-demand agency staff), the uplift in service quality, increased throughput at facilities, and freeing up shift managers’ time.
The process is now also fully automated, whereas previously human operators would laboriously have to collate data from multiple sources (including Excel spreadsheets), which was costly (human resources) and did not have a level of automated quality assurance.
The model is 25% more accurate at shift volume forecasting than traditional human approaches.
Uplift in repeatability and consistency of labor forecasting for planning. We now have a consistent standard and process that can be scaled out nationally, in a consistent and repeatable manner.
Value Brought by Dataiku:
The specific value brought by the Dataiku platform and the Dataiku team include:
Ability to develop, deploy and operationalize ML models at speed and at scale, and within a controlled and governed end-to-end data science workflow.
Dataiku is a single end-to-end integrated data science platform, from development to deployment to BAU operation. This results in a streamlined and consistent process for ML and data science work, across the full spectrum of data science work. Specifically:
The full data science and data engineering lifecycle are native to DSS.
ModelOps and MLOps frameworks are native to DSS, including versioning and dependency management (two key challenges when it comes to production-grade deployments).
The option to leverage advanced models easily and to deploy at scale (using Kubernetes), subject to best-practice MLOps practices as dictated/governed by the Dataiku platform. In the future, we are also looking to leverage Apache Spark as an execution engine within Dataiku as we continue to scale up and roll out the solution nationally.
The ability to leverage Kubernetes to train models at scale, and to easily deploy many models in a production environment, via elastic compute options. Dataiku acts as a seamless abstraction layer from the complexity of the underlying big data processing technologies.
Dataiku enabled us to test many multiple models in parallel, including champion-challenger frameworks, which accelerated the model development and model field testing cycles.
The Dataiku academy and the Australian and global Dataiku teams provided excellent support to uplift our team, and also to support our end-to-end journey from onboarding the platform all the way to our first production deployments and operations, and beyond.