Solar power is expanding rapidly, but the sun remains one of the grid’s most unpredictable suppliers. Clouds, seasonal shifts and changing atmospheric conditions can cause solar generation to rise or fall within hours, creating a difficult balancing problem for utilities. Now, researchers at North Carolina State University have shown that combining several machine-learning models can improve day-ahead solar forecasts by as much as 13% compared with the most consistently performing individual model.
The study, published in the Journal of Cleaner Production, examined how artificial intelligence can predict the amount of solar electricity that will be available roughly one day in advance. Such forecasts are essential for utilities and grid operators, which must schedule power plants, manage energy storage and prepare for fluctuations in electricity demand. As solar energy becomes a larger part of the energy mix, even modest forecasting errors can create operational challenges and increase the need for backup generation.
“Solar power generation has expanded rapidly because it is both renewable and widely available,” said Yen-Hsi Chou, a postdoctoral research scholar at NC State and the study’s corresponding author. “But increasing use of solar can also make forecasting supply and demand more challenging because the availability of sunlight isn’t always consistent.” The researchers focused on the relationship between weather conditions and actual solar power production, using historical observations to train and evaluate a series of predictive models.
The team analyzed operational data collected between January 2019 and December 2022 from two California utilities: the Imperial Irrigation District, or IID, and the Los Angeles Department of Water and Power, known as LADWP. The dataset contained more than 20,000 hours of solar generation and weather information. By studying two geographically and operationally distinct regions, the researchers were able to test whether a forecasting strategy that worked well in one location would also perform reliably elsewhere.
The researchers initially compared seven models drawn from two broad categories: statistical forecasting methods and artificial neural networks. Statistical models are designed to identify recurring patterns in historical data, while neural networks can capture complex, nonlinear relationships between variables over time. This distinction is particularly important for solar forecasting, because the effect of weather conditions on electricity generation is rarely simple. A small change in cloud cover, temperature or atmospheric conditions can produce a disproportionate change in power output.
Among the individual models, a bidirectional long short-term memory network, or BiLSTM, delivered the most consistently accurate results. LSTM networks are a type of recurrent neural network designed to process sequences, making them useful for time-dependent problems such as weather and energy forecasting. A BiLSTM examines information in both forward and backward directions within a sequence, allowing it to identify patterns that may depend on relationships across different time steps. The researchers selected this model as the baseline against which their combined approaches were measured.
The first ensemble strategy used weighted averaging. In this approach, forecasts generated by separately trained, location-specific models were combined, but stronger-performing models received greater influence in the final prediction. This method resembles a panel of experts in which more reliable forecasters are given greater weight. For the IID case, weighted averaging produced the strongest results, improving forecast performance by up to approximately 11% during favorable seasons compared with the BiLSTM baseline.
The second strategy, called a multi-input ensemble, supplied individual models with weather information from multiple locations. Rather than relying only on conditions observed near a particular solar generation area, the system could use broader regional information to improve its understanding of incoming weather patterns. This approach was especially effective for LADWP, where it produced improvements of up to about 13%. The result suggests that meteorological information from surrounding areas may help machine-learning systems anticipate changes that have not yet reached the generation site.
The regional contrast was one of the study’s most important findings. Neither ensemble method performed best everywhere, and no single model consistently dominated across all seasons and locations. “There is no universal forecasting strategy that will perform equally well everywhere,” said Anderson De Queiroz, an associate professor at NC State and co-author of the paper. He added that regional characteristics and the availability of meteorological data must be considered when designing tools for real-world grid operations. A model optimized for one utility may therefore require substantial adjustment before it can be deployed elsewhere.
The findings do not suggest that artificial intelligence can eliminate uncertainty from solar generation, but they show that combining models can make forecasts more robust. Better day-ahead predictions could help utilities decide when to charge batteries, schedule conventional generators and coordinate electricity purchases. The researchers emphasized that ensemble methods must be tested and fine-tuned for the specific region in which they will be used. As solar penetration continues to rise, geographically aware forecasting systems could become an important part of keeping power systems reliable while accommodating more renewable energy.
Subject of Research: Not applicable
Article Title: “A Hybrid Machine Learning Framework for Enhanced Day-Ahead Solar Forecasting in Large Scale Systems”
Web References: https://doi.org/10.1016/j.jclepro.2026.148980; Journal of Cleaner Production article
References: Chou, Yen-Hsi; Haldar, Arundhuti; Nisar, Shubh; de Queiroz, Anderson Rodrigo. “A Hybrid Machine Learning Framework for Enhanced Day-Ahead Solar Forecasting in Large Scale Systems.” Journal of Cleaner Production, published July 29, 2026. DOI: 10.1016/j.jclepro.2026.148980
Keywords
Solar forecasting, machine learning, artificial neural networks, BiLSTM, renewable energy, solar power, ensemble models, weather data, smart grids, energy systems
