Machine learning improves weather predictions for harvesting renewable energy

One of the biggest downsides to renewable power sources like wind and solar is that they are inherently transient. Renewable-backed electrical grids face a unique problem in predicting the weather minutes, hours, and even days in advance. If the error in forecasting is large, it can lead to utilities over- or underestimating their power needs and incurring extra costs.

Sun et al. used machine learning techniques to estimate how frequently short-term forecasts will have extreme errors. Their model is trained on historical weather and energy demand data and is used to reduce the uncertainty in estimates 15 minutes in advance. This method could result in cost savings of up to $20 million per year if implemented.

The researchers designed their system to work within California’s electrical grid, which serves as an ideal testing ground due to its heavy reliance on renewables. They would like to adjust their model to operate in other regions of the country by making more long-term predictions as much as a day ahead. These long-term predictions are especially important for grids that still use coal-based power plants, which require that much time to be brought online.

The authors believe this technique can be applied to make sense of mountains of available data.

“I think with machine learning and power systems, this is just the tip of the iceberg,” said author James Nelson. “With millions of variables and correlations that are very hard to capture via traditional statistics, machine learning can provide some intuition to sort through all of it.”

Source: “Machine learning derived dynamic operating reserve requirements in high-renewable power systems,” by Yuchi Sun, James Henry Nelson, John Colby Stevens, Adrian Au, Vignesh Venugopal, Charles Gulian, Saamrat Kasina, Patrick O’Neill, Mengyao Yuan, and Arne Olson, Journal of Renewable and Sustainable Energy (2022). The article can be accessed at https://doi.org/10.1063/5.0087144 .