Power grid analytics – the data layers that support decisions

The growth in renewables and natural gas in the past decade on the North American power grid has been caused by a significant reduction in levelized costs, tax credits, and a growing demand for cleaner power. Monetization of these assets is critical for investors to realize their internal rates of return (IRR), and how they are traded in deregulated markets is typically the second largest part of that return, behind the prices to which they’re exposed.

The universally agreed-upon benchmark for power trading performance is a percentage of perfect foresight. What is the profit-and-loss for an asset when the future price is already known? When predictions are perfect, the decisions are trivial. It follows that any successful optimization will depend largely on the accuracy of the predictions being used. How then, should power markets analytics teams be evaluating the data they use to optimize their assets?

The three layers: inputs, predictions and decisions

Inputs are known data points about the past and are deterministic (single points). These are measured at weather stations, transformers, and on the ledgers of exchanges and ISOs. They do not exist in the future. Predictions have slowly been commercialized and sold to power markets analytics teams over the past several decades, starting most notably with the GFS and ECWMF weather models in the 1980s. They can be probabilistic or deterministic. Decisions are the financial, operational and ideally optimal actions that are taken based on the best available predictions.  

Each layer is dependent on the one before it – decisions that are made by power traders are a function of how well they can interpret various predictions and what they mean for financial or operational performance.  The strength of predictions are a function of how well a model can match patterns from the past, and the breadth of input data available to it. Inputs are more widely available now with the growth of digitization and sensor deployment worldwide, though satellite imagery and smart metres continue to improve.

In the 1980s, when weather models added weather station inputs, and computational models started matching those inputs to better predictions, power grids benefitted from better decisions and became more reliable.

In-house vs. Outsourcing

Inputs are the first step for any power markets analytics team to solve. Independent system operators, utilities and other government agencies publish power grid meter data. The National Weather Service (NWS) publishes historical weather observations. Commercialization of the input layer is broad, with most analytics teams opting to purchase input data from third-parties.

The prediction layer has been increasingly commercially available. There are very few power market companies that run their own weather models, even though it’s the foundational prediction upon which all other power market predictions are based. Load and renewable generation forecasts built on top of weather models and other inputs are often purchased from third party providers. Price predictions are increasingly available.

The decision layer has significant financial implications for an asset’s profit-and-loss, which is typically why it’s the layer most companies prefer to keep in-house. The decision layer for a load-serving entity may include the volume of load to purchase in the day-ahead market based on existing predictions for load, price and risk on a given hour and with a given commercial hedge already in place. All predictions may be outsourced, or they may generate them in-house.

Every company will have a different approach to how much of the data layer is brought in-house or outsourced. Financial developers may want to outsource all three layers with an off-take, profit share or tolling agreement. Technologically savvy hedge funds may have proprietary input data like private weather balloons that drives their downstream layers. Most power market analytics teams will be somewhere in the middle.

Conclusion

Trends in digitization and sensor availability continue to improve data inputs. More advanced machine-learning applied widely to weather, load and price bolsters models’ ability to predict the future based on historical patterns. Both layers contribute to better decision-making in the final layer, helping asset owners come closer to the widely accepted goal of maximizing their revenue as a percentage of the perfect foresight.

Successful power plant monetization is dependent on the prices to which an asset is exposed, and the decisions that are made using the best available data. Each decision is based on the predictions and inputs available, and the trade-off between outsourcing and developing capabilities in-house will vary depending on a company’s strategy, resources, and risk appetite.

Enertel AI provides short-term energy and ancillary price forecasts for utilities, independent power producers (IPPs), and asset developers to inform trading strategies. We currently operate on the prediction layer. You can view our data catalog, view a clickthrough demo of our product for operators, or request a sample of backcast data for your assets.