Alinta Energy engaged the Ignite team to enable them to demonstrate the capability of their in-house data platform and data science expertise by developing a challenger model to run parallel to their existing gas demand forecast model. This would provide the organisation with the capability to run comparison scenario analysis and accuracy testing of their fundamental business assumptions.
The team delivered:
- Two independent gas demand models for individual states, that can be run concurrently on Databricks.
- Forecast results that can be extracted in a format that is digestible into existing data platforms.
- Forecast versus actual comparisons that evaluate historical performance and optimise forecast accuracy.
The model:
- Uses time series statistics and modelling techniques to create forecasts for multiple geographical areas.
- Enables demand sensitivity insights and risk analysis by incorporating exogenous factors for multiple scenarios.
- Aims to limit errors from simultaneous geographical training.
- Delivers a 5-year forecast horizon across various time intervals and across alternate geographies.
We sat down with Mitch, one of Ignite’s utility data experts and Solution Architects, to garner learnings and perspective from a recent Proof of Value engagement with Alinta Energy. Mitch shares valuable insights about adopting the Challenger versus Champion approach, understanding applicable models, the importance of the Principal Consultant, and the benefits of using Databricks for testing and managing machine learning techniques.
Can you tell us about the Challenger versus Champion approach and how it contributed to the success of this Proof of Value?
Mitch: The Challenger versus Champion approach is all about regularly reevaluating existing ways of working within an organisation. It’s about trialling new tools against existing ones to see which has greater functionality or accuracy. This approach enables organisations to test the accuracy and validity of their existing tools, providing the option to incorporate new assumptions and remove redundant ones in an agile manner. Challenging existing tools and philosophies can create tension, and it’s crucial to ensure that an organisation’s thinking doesn’t become brittle. This approach was at the core of Alinta’s approach to this Proof of Value and played a significant role in its success.
Can you elaborate on the importance of developing a clear understanding of applicable models right from the beginning?
Mitch: When embarking on a data-driven project, it’s essential to strike a balance between delivery management and expectation management. Instead of merely building something that does the job, we focused on building something that successfully proves value. To achieve this, we invested significant time upfront in understanding the business requirements and the data involved. Moreover, we ensured that we had a clear understanding of the different machine learning models applicable to the project. For time series forecasting, there are various options like SARIMAX, Tree Models, LSTMs, RNNs, and Facebook’s Prophet. Choosing the right model should be driven by the initial project requirements, and in our case, Alinta’s had two specific needs that led us down a particular path. Firstly, explainability; when you’re challenging an existing process you need to explain why an algorithm can do it better. Secondly, decomposability; decomposition allows more accurate time analysis, which was a key requirement of the business.
Could you shed some light on the role of the Principal Consultant and their significance in a Proof of Value?
Mitch: Absolutely. The Principal Consultant plays a vital role in a Proof of Value. They are responsible for testing and questioning project requirements to uncover the underlying reasons as to why something is required. This process helps identify the optimum solution and increases the chances of successful adoption of the Data Science model by the business. In our project, we aimed to decouple the business from some central assumptions deliberately. This journey involved testing beliefs, behaviours, processes, and concepts. The Principal’s involvement ensures that the relationship between the technical model and the business use is well-managed. It’s crucial to focus on the business drivers and align strategic intent with operational requirements to avoid complexities and achieve a successful Proof of Value that will ultimately be adopted.
We’re seeing Databricks increase its prominence as the go-to platform for utilities organisations. Why do you think this is?
Mitch: Databricks is an incredible platform for testing machine learning (ML) techniques. It offers the flexibility to test various ML models and select the best one for what is important. Additionally, Databricks provides a robust framework for managing the ML model and its lifecycle, which is incredibly beneficial for developing and continuously improving demand forecasting models. Organisations can automate model training, store models in a registry, and compare them, side by side, against newer versions. This flexibility allows easy access to the model by anyone and facilitates a seamless transition into production. With demand forecasting models relying on diverse data inputs, Databricks enables scalability and sophistication while maintaining accuracy. It also differentiates ML engineers and ML users, allowing users to query at an endpoint, send it data, and return with a prediction.
It’s an all-in-one solution, particularly suitable for large organisations dealing with vast amounts of data and multiple users, as it effectively handles the data lifecycle.
