As solar energy becomes a larger part of the electricity supply, accurately predicting how much power solar panels will generate has become increasingly important.
Researchers at North Carolina State University have now developed a new forecasting approach that can improve day-ahead solar power predictions by up to 13%, helping utilities better manage the electricity grid.
Solar energy is clean, renewable, and widely available, but it has one major limitation: sunshine changes constantly because of clouds, weather, and seasonal conditions.
Since utilities must balance electricity supply and demand every day, they need reliable forecasts of how much solar power will be available in advance.
To improve these predictions, the research team compared several types of forecasting models that use weather data and historical power generation records.
Some of the models relied on traditional statistical methods, which are good at identifying patterns from past data.
Others used artificial neural networks, a form of artificial intelligence that is better at recognizing complex relationships and changes over time.
The researchers tested seven forecasting models using weather and solar power data collected between 2019 and 2022 from two California utilities: the Imperial Irrigation District and the Los Angeles Department of Water and Power.
One of the study’s biggest findings was that there was no single model that consistently performed best in every situation.
Although a model called BiLSTM delivered the most reliable overall performance, the researchers discovered they could achieve even better results by combining predictions from several different models.
The team tested two different methods for combining forecasts.
The first, called weighted averaging, gives greater influence to models that have performed better in the past. The second, known as a multi-input approach, allows the forecasting system to use weather information from multiple locations at the same time.
Both methods improved forecasting accuracy, but their success depended on the region being studied.
For the Imperial Irrigation District, weighted averaging improved forecast accuracy by up to 11%. For the Los Angeles Department of Water and Power, the multi-input method performed even better, improving predictions by as much as 13%.
The findings highlight an important lesson for the growing renewable energy industry. Instead of relying on a single forecasting model for every location, utilities may achieve better results by tailoring forecasting methods to local weather patterns and regional conditions.
According to the researchers, understanding the unique characteristics of each region and combining information from multiple forecasting models can produce more dependable predictions for real-world power systems.
As more homes, businesses, and power companies adopt solar energy, better forecasting tools could make it easier to balance electricity supply and demand, improve grid reliability, reduce operating costs, and support the continued expansion of renewable energy.
The researchers say their results show that combining multiple machine learning models offers a promising path toward more accurate solar forecasting, although each approach should be carefully tested and optimized for the region where it will be used.
