NC State Researchers Achieve Up to 13% Improvement in Day-Ahead Solar Forecasting Using Ensemble Methods
Key Takeaways
- •The study tested seven forecasting models drawn from statistical and artificial neural network approaches using data from 2019 to 2022.
- •The models were evaluated with weather and solar generation data from the Imperial Irrigation District and the Los Angeles Department of Water and Power.
- •BiLSTM was used as the baseline because it was the most consistently strong individual model, but no single model performed best in every case.
- •Weighted averaging improved forecasts by up to 11% for Imperial Irrigation District, while the multi-input method improved forecasts by up to 13% for Los Angeles Department of Water and Power.
- •The researchers said forecasting methods must be adapted to regional conditions rather than applied as a universal solution.
Researchers at North Carolina State University have demonstrated techniques capable of improving day-ahead solar power forecasts by up to 13% compared to the most consistently performing individual forecasting model. Day-ahead forecasts are particularly critical for grid operators, as they inform market scheduling and unit commitment decisions — the process by which power plants are designated to run to meet anticipated demand — made 24 hours in advance. The findings, published in the Journal of Cleaner Production, also underscore the critical role that regional variables play in the design of accurate forecast models.
"As solar penetration continues to increase, forecasting uncertainty becomes a key consideration for energy planners and grid operators that need to balance supply and demand," said Anderson De Queiroz, associate professor of civil, construction and environmental engineering at NC State and co-author of the paper. "Our results demonstrate that combining multiple machine learning-based models can provide more robust predictions and help improve the reliability of solar integration into power systems."
To establish a baseline, the research team first evaluated two broad categories of forecasting models: statistical models, which identify historical patterns, and artificial neural network-based approaches, which are better suited to capturing nonlinear temporal relationships. From these categories, the team selected seven models for testing.
The models were evaluated using weather and solar power generation data spanning 2019 to 2022, sourced from two California utilities: the Imperial Irrigation District (IID) and the Los Angeles Department of Water and Power (LADWP). California leads the United States in installed solar capacity, making it a particularly relevant environment for studying forecasting challenges associated with high solar penetration.
"We wanted to use the models to find the relationship between weather data and solar power generation," said Yen-Hsi Chou, a postdoctoral research scholar at NC State and corresponding author of the paper. "But the most interesting result was that no single model performed best in every case. We chose the most consistently performing model, BiLSTM, as a baseline and saw that by combining the forecasts from different individual models, we could further improve forecasting performance by over 10% in some cases, which was pretty impressive."
The researchers tested two ensemble approaches. The first, weighted averaging, combines forecasts from separately trained, location-specific models and assigns greater weight to better-performing models. The second, known as multi-input, allows each model to incorporate weather data from multiple geographic locations.
Results varied significantly by region. Weighted averaging delivered improvements of up to 11% for IID, while the multi-input approach yielded forecasting improvements of up to 13% for LADWP.
"One of the key takeaways from this work is that there is no universal forecasting strategy that will perform equally well everywhere," De Queiroz said. "Understanding the characteristics of each region and leveraging information from multiple models and locations is extremely important for developing forecasting tools that are both accurate and useful to support decision-making associated with real-world grid operations."
"Overall, our findings indicate that ensemble methods do have the potential to further enhance predictions compared to an individual model, but the methods need to be tested and fine-tuned for the specific region where they will be used," Chou added.
The research was supported in part by the National Science Foundation. Additional NC State contributors to the paper include Arundhuti Haldar, a Ph.D. student in the Department of Electrical and Computer Engineering, and Shubh Nisar, a former graduate student in computer science.
The full paper is available in the Journal of Cleaner Production via ScienceDirect.
