Researchers at North Carolina State University have demonstrated techniques that improved day-ahead solar power forecasting by up to 13% compared with the most consistently performing individual model.

The findings were published in the Journal of Cleaner Production in a paper titled “A Hybrid Machine Learning Framework for Enhanced Day-Ahead Solar Forecasting in Large Scale Systems.”

The research showed that combining multiple machine-learning models can yield more robust predictions and enhance the reliability of solar integration into power systems.

Yen-Hsi Chou, a postdoctoral research scholar at NC State and corresponding author of the paper, said no single model performed best in every case. Combining forecasts from different individual models improved forecasting performance by more than 10% in some cases.

The two ensemble approaches were weighted averaging, which combines forecasts from separately trained location-specific models and gives greater weight to better-performing models; and multi-input, which allows each model to use weather data from multiple locations.

The results showed that forecasting performance depended on regional characteristics and season favourability. Weighted averaging improved forecasting accuracy by up to 11% for the Imperial Irrigation District dataset during favorable seasons, while the multi-input approach improved forecasts by up to 13% for the Los Angeles Department of Water and Power dataset during favorable seasons.

The researchers selected the bidirectional long short-term memory model as the baseline because it delivered the most consistent performance among the individual models.

The models were tested using weather and power generation data collected from 2019 to 2022 from two California utilities: Imperial Irrigation District (IID) and the Los Angeles Department of Water and Power.

The researchers concluded that no single forecasting strategy performs equally well across all regions. They said ensemble forecasting methods have the potential to improve prediction accuracy compared to individual models, provided the methods are tested and tuned for the regions where they will be deployed.

Chou said ensemble methods could outperform an individual model, but their effectiveness must be tested and fine-tuned for the region where they will be deployed.

Anderson De Queiroz, an associate professor at NC State and a co-author of the paper, said understanding regional characteristics and using information from multiple models and locations were critical to developing accurate forecasting tools for real-world grid operations.

Accurate forecasting is becoming increasingly important for power-system planning and operations as solar accounts for a growing share of electricity generation. To reduce deviations in power dispatch through improved forecasting, the Central Electricity Authority in India proposed installing one automatic weather station for each renewable energy project with a capacity of 50 MW.

According to Solargis, a satellite-based approach to monitoring irradiance allows for quality control and error detection and helps calculate the performance ratio. Another smart resource investment in studying irradiance would be pyranometers which give information about solar irradiance in a particular geography by measuring the flux density in solar radiation.