A team of researchers has developed a domain adjustment framework that is able to transfer knowledge of solar energy plants with abundant data to plants that must be trained without labeled data. The framework has been tested on three solar energy sites in Germany and turned out to be better performing than reference models.
Researchers from German Constructor University have developed a new non -controlled domain adjustment framework for predictions for solar energy.
Their technology learns transferable functions of one tanning factory with abundant data and transfers this knowledge to another solar energy plant where the labeled data is absent. It was presented in the newspaper Non-controlled domain adjustment framework for photovoltaic power forecast using variable car encodersPublished in Applied energy.
Corresponding author Atit Bashyal told PV -Magazine That, in contrast to traditional guided approaches, which depend on historical energy data of all sites, “our framework may make accurate short -term prediction possible, even for newly installed PV systems or systems without installed sensors.”
“We achieve this by coordinating the source and target data divisions based on variation-auto-encoders (VAE)-based adjustment, so that the model can effectively generalize about different PV systems without the need for the daisy of the prediction goal,” Bashyal explained.
The research team has called their new architecture the Deep Reconstruction prediction network (DRFN). The DRFN is first trained on a PV factory in a source with a lot of data, at what point it learns how the solar power can predict and reconstruct inputs using VAE.
The model then adapts its power to a new PV factory, which has no data labels. It keeps the predictor for the source and only trains the encoder decoder by minimizing the divergence of Kullback-Libler (KL). KL Divergence is a statistical measure that represents the distance between two distributions.
The architecture was demonstrated on three PV plants in Germany. A solar factory of 1.1 MW was used as a source plant and a 5.8 MW and a 2.5 MW plant were both used as goals. The 5.8 MW plant is 8 km from the source plant, while the 2.5 MW factory is 600 m away.
When predicting the operation of the 5.8 MW factory, the new method had a root average quadratic error (RMSE) of 718.8 kWh, an average absolute error (Mae) of 393.74 kWh and a determination coefficient (R2) of 79.82%. The 2.5 MW target plant achieved 146.78 kWh, 78.94 kWh and 80.49%respectively.
Image: Constructor University, Applied Energy, CC by 4.0
The results were compared with the Smart Persistence prediction method, where a SKILL Index (FSI) forecast was reached for the first goal PV factory and 20.13% for the second. In the case of the first factory, the smart persistence method RMSE of 875.85 kWh, Mae of 466.17 kWh and R2 of 66.3%reached. In the case of the second factory, the results were 183.78 kWh, 98.83 kWh and 65.93%respectively.
Bashyal said that one of the most striking findings was the robustness and effectiveness of the domain adjustment methods in the light of missing land truth data in the target domain.
“Our approach consistently exceeded the smart persistence model and baseline models, despite the absence of target labels during training,” Bashyal explained. “The ablation study further confirmed that our architectural designs have contributed considerably to performance buyers (close to training with data), with the emphasis on the practical potential of our framework in Real-World, data scarce settings.”
Bashyal added that the research team is planning follow -up studies that focus on scaling up the study, integrating incremental or continuous learning and joint modeling of uncertainty in predicting. “These efforts are intended to refine robustness, increase applicability and promote the state of PV prediction about developing implementations and environmental conditions,” he said.
This content is protected by copyright and may not be reused. If you want to work with us and reuse part of our content, please contact: editors@pv-magazine.com.
Popular content
