Researchers at the US Department of Energy’s National Renewable Energy Laboratory (NREL) have enhanced the resolution of wind velocity and solar irradiance data using adversarial training.
The technique was used to improve the resolution of wind velocity data by a factor of 50 and the resolution of solar irradiance data by a factor of 25. This is an unprecedented achievement in climate modelling.
High-resolution climate forecasts are essential for predicting variations in weather harnessed for renewable energy, such as wind, rain, sunlight and sea currents. Short-term forecasts are essential for operational decision-making, while longer-term climate forecasts are useful for scheduling, resource allocation and infrastructure planning.
However, it is tricky to preserve temporal and spatial quality in climate forecasts, with the lack of high-resolution climate data proving a serious challenge for energy resilience planning.
“To be able to enhance the spatial and temporal resolution of climate forecasts hugely impacts not only energy planning, but agriculture, transportation and so much more,” said NREL computer scientist Ryan King.
Although various machine learning techniques have emerged to enhance data through super-resolution (the process of sharpening fuzzy images by adding pixels), no one had previously used adversarial training for this purpose.
The scientists used this machine learning technique to train a network to produce physically realistic details by observing entire fields at a time. The adversarial training involved making neural networks compete with each other to generate more realistic models; one network picked up patterns in existing solar irradiance and wind velocity data, while the other inserted these characteristics into coarse data, sharpening its resolution.
Over time, the networks became far more effective at producing realistic data, as well as distinguishing between fake and real inputs. The researchers were able to add 2,500 pixels for every original pixel.
“Adversarial training is key to this breakthrough,” commented Andrew Glaws, a machine learning specialist and co-author of the Proceedings of the National Academy of Sciences study describing the new approach.
This method will enable scientists to complete renewable energy studies in future climate scenarios with greater speed and accuracy, the researchers said. It is applicable to a range of climate scenarios, from regional to global scales.