Team members:
- Yu Chen, Data Scientist
- Mrig Debsarma, Data Science Manager
- Stevani Kho, Business Strategist
- Benn Martens, Energy Trading Analyst
- Jawad Hassan, Data Engineer
- Megha Agarwal, Data Engineer
- Guillermo Godoy, Electricity Risk Manager
- James Holgate, Head of Data and Analytics
Country: Australia
Organization: Ampol Limited
Ampol supplies fuel to approximately 80,000 business customers in diverse markets across the Australian economy, including defense, mining, transport, marine, agriculture, aviation and other commercial sectors. Our retail network also serves approximately three million customers every week with fuel and convenience products.
We now also have the goal of becoming the supplier of choice with alternative energy solutions to power better journeys including the development of electric vehicles charging stations, retail electricity offering, and other non-traditional transport fuels present an opportunity for Ampol to decarbonize the transport market and leverage our existing position into clean charging infrastructure.
Awards Categories:
- Best Moonshot Use Case
- Best MLOps Use Case
Business Challenge:
Energy Transition and Decarbonization is one of Ampol’s strategic pillars, a key priority area that will help ensure we deliver on our purpose to ‘power better journeys today and tomorrow’. In addition to continuing to supply fuel to customers in diverse markets, Ampol have the goal of becoming the supplier of choice with alternative energy solutions, including the launch of a retail energy business Ampol Energy and electric vehicle charging brand AmpCharge.
As Ampol embarks on building this new electricity business, robust processes are needed to support participation in the market. The Australian electricity wholesale market is highly dynamic, with settlements every 5 minutes and potential for prices to vary significantly across 5-minute periods. There are large structural differences between electricity wholesale market where electricity is bought and the retail markets where it is sold. As a result, there is a significant financial exposure risk which needs to be managed in both the short and medium term.
Granular and up-to-date predictions of electricity load are key to enabling exposure management for the business and informs risk valuation and exposure management strategies. “Without a load forecast, there is no electricity business” says the Electricity Risk Manager.
A significant challenge to the Load Forecasting project is the desired speed of market entry, leading to the need to deliver modelling and forecasts with limited training data. When the project commenced:
- There were no retail energy customers supplied by Ampol.
- The electric vehicle charging network was still in early stages of rollout.
- There was no existing process for demand management.
Our data science team leaned into the challenge to build a new solution from scratch which supported a successful wholesale electricity market entry
Business Solution:
Ampol needed to leverage data, machine learning, and risk management expertise in order to establish a fit-for-purpose load forecasting capability and processes.
A cross-functional team was established, including business stakeholders who collaborated with the Data and Analytics team to define the forecasting project objectives and approach to be used. Once agreed, we proceeded with an end-to-end ML delivery process.
Dataiku was used as a comprehensive environment to deliver this process, including:
- Data ingestion of new external sources using out-of-the box Dataiku API Connectors as well as historical file-based ingestions
- Data preparation including cleaning raw data, transforming JSON data files, joining datasets, and other data enrichment
- Data pipelines, which were set up and scheduled to provide ongoing success of jobs.
- Exploratory Data Analysis (EDA) conducted to understand the dataset and validate modelling hypotheses
- Model development, carried out by data scientists
- Sharing of results and iteration of the models with key stakeholders
- Productionisation of models once development was completed, through the bundling and migration of the project to a DSS Automation instance
- Implementing metrics and checks to establish quality assurance
- Automation of the model execution process
- Establishing the forecast lookback process to continuously improve the model
Day-to-day Change:
This work has resulted in the establishment of an end-to-end Load Forecasting business process, which is a key requirement for energy market participation and financial risk management.
The Load Forecasting output is used across many business activities with applications include:
- Informing traders to develop contracting and hedging strategies
- Informing retail pricing
- Informing prudential considerations related to electricity risks
- Understanding the demand drivers in the market and generating market insights
Ultimately the Load Forecast work enables data-driven decisions to be made in all the subsequent processes it touches and gives the business confidence that we have a timely view of the market accounting for key factors.
The rapid delivery of this use case, and the associated data insights, has supported the establishment of strong analytical decision making and data-driven culture in a new business area. Our ability to collaborate with business users and deliver insights has built confidence in the team for ongoing growth of this internal analytics capability.
Business Area Enhanced: Risk/Compliance/Legal/Internal Audit
Use Case Stage: In Production
Value Generated:
The accurate electricity load forecasting provided by Dataiku has resulted in improved decision-making for Ampol Australia's Wholesale Energy trading team.
The Load Forecast enables effective management of millions of dollars of potential financial exposure in the electricity business and is a key enabler for Ampol to continue to roll out home retail electricity and electric vehicle charging solutions for customers. Without this process the financial risk of entering these new markets would be significantly higher and could delay the implementation of Ampol’s energy and decarbonisation strategy.
Agile development of the modelling, leveraging Dataiku, has resulted in a fit-for-purpose solution achieved in the necessary timeline and a low total cost to the business.
The capability established through this initial use case provides a foundation on which to deliver additional high value use cases in these energy markets in the future.
Value Brought by Dataiku:
Dataiku is a comprehensive platform which facilitated effective delivery of the project, by a team of different users, ultimately resulting in better quality solutions and accelerated project timelines.
A key benefit enabled through the platform is the platform accessibility and model explainability features. Dataiku enables both technical managers and business stakeholders to see the pieces of the puzzle and how they fit together, an out-of-the box functionality to understand how the models work. As a result, the end solution delivered in Dataiku never needs to be a ‘black box’. This has been fundamental in building confidence in the models and how they work with diverse stakeholders – from leaders right through to technical SMEs.
Leveraging Dataiku's advanced tech stack deployed on Azure Kubernetes Service (AKS) means that Ampol has been able to leverage innovative data engineering and data science techniques, at scale, without needing to solve for infrastructure, access, security, or network questions. This extends to model version control and archiving features where production models can be packaged and deployed in a controlled fashion, with extreme ease.
The ability of the platform to set up schedules, define model metrics and implement data checks, means that execution management processes are all in one place. As a result, developers can contribute directly to the needs of MLOps, while keeping their focus on development.
Value Type:
- Improve customer/employee satisfaction
- Increase revenue
- Reduce cost
- Reduce risk
- Save time
Value Range: Dozens of millions of $
