Short-term solar forecast using convolutional neural networks with sky images
Abstract/Contents
- Abstract
- Solar photovoltaic (PV) capacity is rapidly growing across the world. However, the volatility of cloud movement introduces significant uncertainty in short-term solar PV output, which can complicate the operation of modern power systems. Cloudy days remain challenging for modern short-term solar forecasting algorithm. An improved short-term forecast benefits all participants of the sub-hourly power market. This work proposes a specialized convolutional neural network (CNN) "SUNSET" for short-term solar PV output forecasting. Its suitability is first tested on now-casting, i.e. inferring contemporaneous PV output from sky images. On a system with a rated capacity of 30.1 kW, the baseline SUNSET model achieved an RMSE of 1.01 kW on the sunny test set, 3.30 kW on the cloudy test set, and 2.40 kW overall. This validates the sky images' close correlation with PV panel outputs and that a CNN is suitable to extract this correlation. Extensive experiments are done to optimize the structure of SUNSET. In terms of depth, having three convolutional layers and one fully-connected layer produces the best result. Both types of neural nets are found to be crucial for model performance. In terms of width, 48 filters in the convolutional layers and 2048 neurons in the fully-connected layers provide the best performance. In terms of image resolution, 64 x 64 is the optimal point, as either finer or coarser resolution results in worse RMSE. Two further techniques are also found to be useful: drop-out increases the robustness for generalization while ensemble modeling decreases forecast error. For forecast, the SUNSET model is augmented in two key aspects, the usage of hybrid input and temporal history. PV output history is injected mid-way in the model to be joined with the processed image features. The temporal history of sky images are included by concatenating the images in the color channel. On a 1-year database, the "baseline'' model achieves a 15.7% forecast skill in all weather conditions, and a 16.3% forecast skill in the more demanding cloudy conditions, relative to a smart persistence forecast. Optimal input and output configurations for forecast are also explored. In terms of input, both sky images and PV output history are found to be crucial. Output-wise, training against PV output significantly out-performs training against clear sky indices (CSI). Careful down-sampling can reduce the training time by as much as 83% without affecting accuracy. For lag term configurations, using the same length of history as the forecast horizon is a good heuristic, while using slightly shorter history yields a modest 0.5% - 0.9% improvement. Last but not least, a two-stage optimization framework is proposed to quantify the value of short-term solar forecast. Design optimization in the first stage solves for resource capacity, while the receding horizon control (RHC) in the second stage simulates a power system's operation for a year. Within this framework, we consider a microgrid scenario which have battery as an option, and a demand charge scenario which allows grid import. In the settling process of the RHC stage, batteries can be utilized and redesigned to address forecast error in the microgrid scenario, while grid import is incurred in the demand charge scenario for the same purpose. For the microgrid scenario, a perfect forecast can reduce overall cost by 3.6% to 12.5% comparing to a persistence forecast. The largest cost savings are achieved with medium solar penetration of 30% to 60%. For the demand charge scenario, a perfect forecast can reduce total cost by 8.9% to 25.6%. For SUNSET with a forecast skill of 15.7%, we can expect a cost saving of 1-2% or 2-5% respectively in these two scenarios.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2019; ©2019 |
| Publication date | 2019; 2019 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Sun, Yuchi |
|---|---|
| Degree supervisor | Brandt, Adam (Adam R.) |
| Thesis advisor | Brandt, Adam (Adam R.) |
| Thesis advisor | Durlofsky, Louis |
| Thesis advisor | Lobell, David |
| Degree committee member | Durlofsky, Louis |
| Degree committee member | Lobell, David |
| Associated with | Stanford University, Department of Energy Resources Engineering. |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Yuchi Sun. |
|---|---|
| Note | Submitted to the Department of Energy Resources Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2019. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2019 by Yuchi Sun
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...