A team of researchers at the US Department of Energy’s (DOE’s) National Renewal Energy Laboratory (NREL) has created a new machine learning approach to quickly improve the blurry resolution of wind velocity data by 50 times and solar irradiance data by 25 times. This kind of enhancement that has ever been achieved before the climate data.
The team took an alternative approach by using adversarial training. The model produces realistic details by observing entire fields at a time, generating high-resolution climate data at a much faster rate. This new method will allow scientists to complete renewable energy studies in future climatic scenarios with more accuracy.
Ryan King, a senior computational scientist at NREL who specializes in physics-informed deep learning, explained that enhancing the temporal and spatial resolution of these climate forecasts hugely impacts not only the energy planning process but transportation, agriculture, and various other factors.
The NREL team presented a new article detailing their approach, titled ‘Adversarial super-resolution of climatological wind and solar data,’ published in the journal Proceedings of the National Academy of Sciences of the United States of America.
High-resolution and accurate climate forecasts are essential for predicting shifts in wind, rain, clouds, and sea currents that fuel renewable energies. Short-term forecasts drive the decision-making process in operations, medium-tier, weather forecasts, scheduling and resource allocations, and long term climate forecasts to inform infrastructure planning and policymaking.
Still, it is challenging to preserve a temporal and spatial quality in these climate forecasts, the team says. The lack of high-resolution data for different conditions has been a significant challenge in energy resilience planning. Several machine learning techniques have emerged to enhance the coarse data through high resolution—the classic imaging process of sharpening a blurry image by increasing pixels. But until now, no organization has used adversarial training to super-resolved climate data.
Adversarial training essential for improving the performance of neural networks by having them compete with each other to generate more realistic data. The NREL team trained two types of neural networks in the model, one to detect physical characteristics of high-resolution solar irradiance and wind velocity data, and another to insert clearer characteristics into the unedited data. With time, the networks produce more realistic pictures and improve at differentiating between real and fake inputs. The team was able to add 2,500 pixels for every original pixel.
The team also explained that using adversarial training instead of the traditional numerical approach to climate forecasts, which can involve solving many physics equations, saves computing time and data storage costs. It also makes high-resolution climate data more accessible. This new approach can be used in a wide range of climate scenarios from regional to a global scale, shifting the paradigm for climate model forecasting.