From habitats to property to livelihoods, wildfires destroy everything in their path. But there is another, less-acknowledged, casualty: sunlight and the electrical grid that depends on it. Smoke from wildfires can cover large swaths of land, including solar farms, and significantly reduces power production from photovoltaic (PV) panels.
In response, Cornell researchers have created a machine learning-based model that can forecast, with greater accuracy than current methods, the impact severe wildfire conditions will have on solar electricity generation. This will enable system operators to better match supply and demand, and keep costs down.
"If you don't have a good forecast, then you have to rely on your so-called reserve generators, which are very costly," said Max Zhang, the Irving Porter Church Professor of Engineering in Cornell Engineering and Provost's Fellow for Public Engagement, who led the project. "As we have more solar energy penetrating into the power systems, the economic consequence can be higher and higher."
The research was published April 22 in Environmental Research Letters. The paper's co-lead authors are Fenya Bartram '25 and Bo Yuan, M.S. '23, a Ph.D. student in mechanical engineering.
Zhang first recognized the threat to solar energy production in the summer of 2023, when the northeastern U.S. was blanketed in smoke from Canadian wildfires and PV output in the region dipped.
"I got a lot of interview requests regarding the air pollution and health effects," Zhang said, "but I was also wondering, how about the energy side?"
Zhang and his team found that the day-ahead forecasts made by the New York Independent System Operator (NYISO), which monitors and coordinates how the state's power system operates, significantly overpredicted PV output during the wildfires.
"There are day-ahead markets and real-time markets. They need a forecast of the energy production in order to balance supply and demand," Zhang said. "Either overprediction or underprediction is not good, especially overprediction."
The researchers set about building a machine-learning model by incorporating a series of public domain data products from the National Oceanic and Atmospheric Administration's new High-Resolution Rapid Refresh Smoke (HRRR-Smoke) weather forecasting system, which included predictions of aerosol impacts and smoke mass density during severe wildfire periods.
Zhang's team is the first to harness the system's power of prediction for this kind of application. The fact that HRRR-Smoke played such an essential role demonstrates how the public benefits from government climate data tools.
"If we don't have enough people of talent maintaining and improving those products, then that will cause damage to many sectors of society," he said.
One of the factors that makes forecasting wildfire smoke disruptions so difficult in New York state is that the occurrences are so rare - though that could change as climate change exacerbates extreme weather events. To compensate for the current dearth of regional data, the team employed "upsampling" - i.e., increasing the sampling rate - to train their model to put more emphasis on wildfire events, despite their infrequency.
The team tested the model using hourly solar data collected by the New York State Energy Research and Development Authority (NYSERDA) - which supported the research - during previous wildfire periods, and they determined the model outperformed NYSIO's forecasts. While other researchers have been working to better predict power production in the aftermath of the western wildfires, the tool created by Zhang's team is the first to operate on an hourly basis, rather than on daily averages.
"Everything reported in our paper is operational," he said. "All the inputs we use in the model are forecast products. That's what power system operators need. And it can be used anywhere."
Zhang anticipates that increases in solar development, combined with more frequent wildfires, will make forecasting high smoke periods and the impact on solar electricity production even more critical for maintaining the power system reliability in New York state and across the country.
"This is just the start. We are improving the model while creating pathways for adoption by system operators," he said. "The better the forecast, the more reliable the power system."
