By: Andrew Yule
With more data and information available than ever, we can solve previously unsolvable problems using artificial intelligence and machine learning. By creating data-driven predictive models, machine learning unlocks new opportunities to scale up processes that would have previously required human intervention. The best applications of these tools help organizations become more efficient and effective in their day-to-day operations.
However, while machine learning and AI can solve problems once deemed too complicated, they have their limitations. So, while it’s important to know when to use AI and machine learning in the energy industry, it’s just as important—if not more important—to know when not to use it.
When’s The Right Time To Use Artificial Intelligence And Machine Learning In The Energy Industry?
When used appropriately, machine learning and AI can help solve many problems and provide improved insights that a pure physics-based model can’t capture. These technologies can automate redundant work, improve accuracy and efficiency, and create new opportunities by augmenting or replacing traditional modeling methods.
AI and machine learning can also have positive implications and impacts on Environmental, Social, and Governance efforts. ESG is highly data-intensive, and many organizations struggle to extract actionable insight from all this information. Applying these technologies to large data sets can result in insights that support leak and restriction detection, flare detection, CO2 transport, reducing emissions, and lowering an organization’s overall carbon footprint.
But, when these tools are overused, they can deliver erroneous or useless results or create costly, unnecessary model maintenance. In some cases, traditional engineering calculations could outperform machine learning models—especially in cases where there is only a limited amount of data.
5 Questions We Ask Before Using Artificial Intelligence Or Machine Learning On Your Project
With our strong engineering roots, access to the latest physics-based software, experience in AI/ML approaches, and focus on deeply understanding and solving our partners’ problems, SPL continuously evaluates when it’s most appropriate to apply which technology.
We use ML to solve problems by implementing a large- or small-scale digital twin model of the process that allows us to experiment with various inputs to predict outcomes freely.
But, before we create that model, we ask critical questions about the project, your organization, and your goals to ensure that we’re using the right tool for the job.
1. Are traditional engineering methods applicable?
The first question we ask our team before pursuing AI or ML is if traditional engineering methods work. Because if they do, there is no need to reinvent the wheel.
However, If a traditional engineering method has a shortcoming, then machine learning models might be able to fill that gap.
The true value of machine learning comes when a high number of variables or mutual parameters affect your results, making a physics-based model unfeasible. Machine learning also becomes more valuable when sufficient historical data is available and can be applied to build a predictive model. Without historical data, traditional engineering models will likely perform better.
2. Does your problem fit a typical use case of machine learning?
If your problem fits in a typical machine learning use case such as predictive maintenance, production forecasting, or modeling for exploration, machine learning is likely a good choice. Machine learning can be beneficial in forecasting and predicting data and trends as the variables involved become more complex.
However, even if your model doesn’t fit in a typical use case for machine learning, it doesn’t mean machine learning can’t be applied.
3. What is the state of your data?
Your organization’s data collection processes, governance, and the accuracy, completeness, consistency, validity, uniqueness, and timeliness of that data are crucial to understanding if using ML or AI is appropriate. For example, if you have organized data that’s automatically selected and stored at regular intervals, it is more reliable and more likely to create a representative machine learning model successfully.
However, if the data is not available in a centralized location or it’s not in suitable condition to be used, some data engineering will be required. If there is not enough quality data, creating models can be difficult. Conversely, an overabundance of data can result in other processing-related challenges.
4. Do you need to be able to explain your model?
ML is a bit of a black box, so if you need to deploy your model, get results, and don’t necessarily need to explain it, the machine learning models can work well. However, if you need to explain the mechanics and workings, then machine learning may not be the best tool to use.
5. Do you need to deploy your model? Does it need to be regularly updated as field conditions change?
Provided you have an underlying physics-based model, deploying such a model is much easier as changes in the system over time are theoretically captured automatically. On the other hand, ML will consistently require updating the model with new data as the system changes in time. Without the proper data architecture in place, this can be very difficult.
So, Is AL/ML Right For Your Project?
After a hyper-growth period, the energy industry is now normalizing its use of machine learning and AI. As the technology matures, there will be incremental improvements in what is possible, resulting in gains for organizations looking to leverage it appropriately. At SPL, we continue to stay current on the best use of these tools to help our partners have accurate, defensible data at their fingertips.