Machine Learning for Power Traders

Domain expertise, clean data and modelling software are key requirements to properly forecast time series data in power markets

At its core, artificial intelligence is about using things we know to predict the unknown. Self-driving cars use data about their surroundings to predict the optimal driving decision, using behaviour from the past. Image recognition uses information about each of the pixels in an image to predict how a human might interpret it, using billions of documented images. In power markets, we can use information from weather stations, gas prices, transmission outages and generation outages to predict power prices, generation and load patterns. There are three key ingredients required to make accurate predictions on a power trading desk: the expertise from a trader to know what data is likely to impact the forecast target, the data itself, and the software required to train, evaluate and use models in production.

Target: A target is a time series the user wants to predict. Typically, it’s a combination of an object on the grid (price node, hub, weather zone) and a type of data associated with that object (load, generation or a price).

Model: A model is a mapping between input variables and target time series that can be learned from historical data.

Domain Expertise

There’s a considerable amount of domain expertise required to intelligently transact in power markets — you wouldn’t want someone without a driver’s license designing your self-driving car. Many power traders have specific analyses they conduct prior to submitting bids and offers, and that reasoning is borne from many years of familiarity with specific power market patterns. A good machine learning model will include those learnings and allow power traders to steer model development by adding their years of expertise to the model inputs.

The Data

There’s very little debate about how much data is required for the best predictions: more is almost always better. Google and Facebook are so good at predicting conversions for their advertisements because they know so much about their users. Self-driving cars are safer if they know about nearby bicycles and pedestrians, and not just the road in front of them. The same is true in power markets: models that have the option to include many inputs tend to outperform models that only include load forecasts.

Modelling Software

The last input required to use AI in power markets is the software that combines the domain expertise of power traders with clean, relevant data and state-of-the-art machine learning algorithms. The software should also be able to run on a schedule and automatically evaluate the accuracy of the models in real time. Technically, Excel spreadsheets could fill this need — and often do. More advanced software organizations may use custom scripts written in R or Python (among other programming languages). But while programming reduces manual tasks, it can require expensive, specialized expertise and comes with significant R&D investment.

All three inputs are required for trading desks to effectively use artificial intelligence in their trading activities. When software engineers build the latest models with fantastic data, but don’t understand how inputs affect the target (for example, by ignoring the price of gas), their predictions can underperform. If a veteran power trader uses an Excel model with clean, reliable data inputs, they could be underperforming compared to someone else who may have access to the latest in AI software and spends less time assessing their model accuracy. Finally, if someone has considerable domain expertise and the latest in software, but no access to high quality and reliable data, their forecast accuracy may suffer and they will be spending much of time on cleaning data.

Enertel AI is a platform built in collaboration with Yes Energy data inputs that allows power traders to design machine learning models without writing code. Users can search from Yes Energy’s vast array of data sources and build machine learning models in less than ten clicks. Interested in adding reserve margins to your model? Gas prices? Available transmission capacity? All are available through Yes Energy and Enertel AI’s partnership. Learn more.