AI for Science
A hybrid solar-power model lowered short-term prediction error in tests at six photovoltaic plants, a 2026 study reports
In a 2026 paper published in Energies, a peer-reviewed journal, researchers report a model that made short-term photovoltaic power predictions more accurate. Photovoltaic power is electricity made from sunlight. Tests used datasets from six photovoltaic plants. The study reported lower mean absolute error, lower root mean square error, and a higher coefficient of determination, a statistic that measures how well predicted values match actual values. Those gains came in comparison with three existing models named Transformer, Reformer, and Informer. The abstract reports the model was especially robust and efficient in difficult conditions such as rainy weather.
What did the researchers do?
The researchers proposed a framework with several machine-learning parts. It begins with an anomaly-detection step, a way to flag unusual data points, called the Artificial Fish Swarm Algorithm-Isolation Forest (AFSA-IF). A generative adversarial network, a machine-learning method with two competing neural networks, performs feature extraction. The framework fuses weather data with past power output and classifies weather conditions before using a Patch Time Series Transformer (PatchTST) model. The study tested this approach on datasets from six photovoltaic plants.
What were the main results?
Compared with Transformer, Reformer, and Informer models, the proposed model showed significant improvements in mean absolute error, root mean square error, and coefficient of determination. Those gains mean the short-term solar-power predictions matched actual output more closely on the tested datasets. The abstract says the results confirm the model's robustness and efficiency, especially in rainy weather.
Frequently asked questions
What is photovoltaic power prediction?
It is forecasting how much electricity a solar photovoltaic plant will produce over a short period ahead. This study combines weather data with past power records to make those short-term forecasts.
Which models did the researchers compare?
The comparison models were Transformer, Reformer, and Informer.
Which challenging weather condition did the study mention?
The abstract gave rainy weather as an example of a challenging condition where the model was robust and efficient.
Limits and what's next
The abstract does not state limitations or next steps.
Source: Xin Lv et al., “Patch Time Series Transformer−Based Short−Term Photovoltaic Power Prediction Enhanced by Artificial Fish,” Energies, 2026.