The search for optimal short-term solar photovoltaic forecasting using neural networks

NOV 15, 2019

Researchers use a convolutional neural network system to explore fusing image data with photovoltaic-generation history in the search for an optimal forecasting model.

Jodi Ackerman Frank

DOI: 10.1063/1.5133372

The search for optimal short-term solar photovoltaic forecasting using neural networks internal name — The search for optimal short-term solar photovoltaic forecasting using neural networks lead image

Although solar power continues to grow in popularity, its intermittency makes the renewable energy source unreliable. The primary problem is cloud cover, which can fluctuate throughout the day, making short-term solar forecasting an essential need.

Venugopal et al. explore the challenges of fusing information from image-based and non-image-based inputs in the search for an efficient solar forecasting model. To do so, they studied 28 methods of fusing these data inputs in a convolutional neural network (CNN) system they built called SUNSET, which incorporates video-camera sky images.

“Most of these fusion methods are inspired from past studies in the field of robotics,” author Vignesh Venugopal said. “The idea is to find the best CNN-based architecture for using photovoltaic (PV) output history and sky image history to predict 15-minute-ahead future PV outputs.”

Study parameters were divided into the fusion methods and the hyperparameters, which control what and how the CNN model learns. Fusion methods were classified as mix-in, activation map combination, activation map stacking and two-step.

The researchers found the best fusion methods were the two-step autoregressive model and the mix-in Yuan-Model. In the first fully connected layer of the Yuan-Model, some neurons have access to the image-based inputs, some have access to the non-graphical inputs and some have access to both. When input data modalities differ significantly, focusing on separate low-level features is effective.

However, the two-step model thrives on using the PV log first before bringing sky images into play so that PV log values aren’t lost in a vast amount of information from the images.

Future work will focus on determining the effects of other weather parameters, such as temperature, and comparing the results with additional analysis methods.

Source: “Short-term solar PV forecasting using computer vision: The search for optimal CNN architectures for incorporating sky images and PV generation history,” by Vignesh Venugopal, Yuchi Sun, and Adam R. Brandt, Journal of Renewable and Sustainable Energy (2019). The article can be accessed at https://doi.org/10.1063/1.5122796 .