How machine learning can boost accuracy of solar power forecasts

As solar energy plays an increasing role in the global power supply, ensuring accurate forecasts of photovoltaic (PV) power generation is critical for balancing energy demand and supply.

A new study published in Advances in Atmospheric Sciences explores how machine learning and statistical techniques can refine these forecasts by correcting errors in weather models.

Weather forecasts are a key input for PV power prediction models, yet they often contain systematic errors that impact accuracy. Researchers from the Institute of Statistics at the Karlsruhe Institute of Technology examined different ways of improving these predictions by applying post-processing techniques at various stages of the forecasting process.

Their study tested three strategies: adjusting weather forecasts before they enter PV models, refining power predictions afterward, and using machine learning to forecast solar power directly from weather data.

“Weather forecasts aren’t perfect, and those errors get carried into solar power predictions,” said Nina Horat, lead author of the study. “By tweaking the forecasts at different stages, we can significantly improve how well we predict solar energy production.”

The findings reveal that post-processing enhances solar power predictions the most when applied to power forecasts rather than weather inputs. While machine learning models generally outperform traditional statistical methods, their advantage in this case was limited—likely due to the available input data. The study also found that including the hour of the day as a factor was crucial for accuracy.

“One of our biggest takeaways was just how important the time of day is,” said Sebastian Lerch, corresponding author of the study. “We saw major improvements when we trained separate models for each hour of the day or fed time directly into the algorithms.”

One promising approach bypasses traditional PV models entirely, using a machine learning algorithm to predict solar power directly from weather data. This method offers a practical advantage: it does not require detailed knowledge of a solar plant’s design, though it does need historical weather and performance data for training.

The research opens the door for future studies to refine machine learning approaches further, integrate additional weather variables, and extend analyses to multiple solar plants.

As renewable energy continues to grow, improving forecasting techniques will be key to ensuring a stable and efficient power grid.