Multifocus Convolutional Neural Networks for Short-term Solar PV Power Output Prediction
Abstract/Contents
- Abstract
Reliable forecasts for the power output from variable renewable energy generators like solar photovoltaic systems are important for balancing the load on real-time electricity markets and ensuring the reliability of the grid. However, solar PV power output is highly uncertain, with significant variations occurring over longer (daily or seasonally) and shorter (minutely) timescales due to weather conditions, especially cloud cover. Therefore, accurate models that can provide these predictions are necessary.
This thesis builds on existing work, the SUNSET model, that uses a convolutional neural network applied to the computer vision tasks of nowcasting and forecasting solar PV power output. Multifocus SUNSET Nowcast and Forecast model architectures are presented and compared to the baseline SUNSET models. The multifocus models utilize a two-stream approach: one branch ingests a 64 x 64 pixel downsampled full sky image and the other branch ingests a 64x64 pixel sun patch cropped from a 256 x 256 pixel resolution full sky image. Various locations of fusion within the model were performed in order to optimize for accuracy as well as for model size and parameter count.
It was found that the best Multifocus SUNEST Nowcast model outperformed the baseline model by 9.8% on the test set when considering images without border masking and by 14.8% when considering images with border masking. This large improvement comes from better performance on cloudy test days. Further, the best Multifocus SUNSET Nowcast model fused the two branches after the fully connected layers but before the regression. However, the multifocus model that fuses after the 1st convolution layer had a similar performance on the validation set with far fewer parameters. Thus, it may be worth utilizing the multifocus model that fuses after the 1st convolution layer instead if computational resources are limited. When considering the best Multifocus SUNSET Forecast model, with respect to the baseline model, only minor improvements in performance were noticed on the order of 0.01 kW measured as RMSE, relative to baseline RMSE of 2.62 kW. Thus, the Multifocus SUNSET Forecast model may not be the best choice for forecasting because it incurs additional computational costs compared to the baseline model without improving performance.
Description
Type of resource | text |
---|---|
Publication date | December 12, 2022 |
Creators/Contributors
Author | Scott, Andea |
---|---|
Advisor | Brandt, Adam |
Subjects
Subject | Solar forecasting |
---|---|
Subject | Convolutional Neural Networks |
Subject | Machine learning |
Subject | Deep learning |
Genre | Text |
Genre | Thesis |
Bibliographic information
Access conditions
- Use and reproduction
- User agrees that, where applicable, content will not be used to identify or to otherwise infringe the privacy or confidentiality rights of individuals. Content distributed via the Stanford Digital Repository may be subject to additional license and use restrictions applied by the depositor.
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 4.0 International license (CC BY-NC).
Preferred citation
- Preferred citation
- Scott, A. (2022). Multifocus Convolutional Neural Networks for Short-term Solar PV Power Output Prediction. Stanford Digital Repository. Available at https://purl.stanford.edu/yq210kc0011. https://doi.org/10.25740/yq210kc0011.
Collection
Master's Theses, Doerr School of Sustainability
View other items in this collection in SearchWorksContact information
- Contact
- andea98@stanford.edu
Also listed in
Loading usage metrics...