The document explores enhancing the accuracy of solar power generation predictions using a generative adversarial network (GAN) and long short-term memory (LSTM) models. It outlines the challenges posed by unpredictable solar power generation and discusses the benefits of data augmentation in improving prediction models. The proposed LSTM model achieved a root mean square error (RMSE) of 0.1898, which was further reduced to 0.0802 with the augmented dataset generated by the GAN.

1. Enhancing the Prediction Accuracy
of Solar Power Generation using
a Generative Adversarial Network
Kundjanasith Thonglek1
, Kohei Ichikawa1
, Keichi Takahashi1
,
Chawanat Nakasan2
, Kazufumi Yuasa3
, Tadatoshi Babasaki3
, Hajimu Iida1
1
Nara Institute of Science and Technology, Nara, Japan
2
Kanazawa University, Ishikawa, Japan
3
NTT FACILITIES, INC., Tokyo, Japan
1

2. Produce stable electricity with solar power generation
Cloud
Wind
A commonly known problem in solar power generation is
Controlling the amount of stored electricity in batteries to produce stable electricity
➢ Unpredictable
➢ No pollution
➢ Predictable
➢ Pollution
2

3. Prediction of solar power generation
➢ There are several existing approaches to tackle controlling the amount of
stored electricity in batteries to produce stable electricity
○ Designing a new energy storage [1]
○ Improving the capacity of the batteries [2]
○ Building a prediction model for the solar power generation [3]
➢ Applying a new energy storage or developing the existing batteries requires
large amounts of investment more than embedding prediction models
○ Therefore, the development of highly accurate prediction models
such as deep learning has received much attentions
References
[1] C. Thombre, S. Shah, M. Mahajan, and T. Haldankar, “Design of a battery-less solar energy storage system based on re-generation of energy”, in Proceedings of IEEE International Conference on
Computing, Communication and Networking Technologies (ICCCNT), 2017, pp. 1-5.
[2] Z. Hradiflex, P. Moldrik, and R. Chvalek, “Solar energy storage using hydrogen technology”, in Proceedings of IEEE International Conference on Environment and Electrical Engineering (EEEIC),
2010, pp. 110-113
[3] V. Prema and U. Rao, “Development of statistical time series models for solar power prediction”, Renewable Energy, Vol 83, 2015, pp. 100-109 3

4. Deep learning approaches
➢ Deep learning is a kind of machine learning approach that applies neural networks
with many layers
○ It has been proved to be highly effective in a wide range of prediction
➢ Applying LSTM is outperformed for predicting solar power generation [1]
○ Solar power generation is a kind of time-series data
○ LSTM
■ It is a type of recurrent neural network
■ It is able to hold long-term historical data
■ It is able to learn contexts from time-series data and forecast future trends
References
[1] J. Zhang, Y. Chi, and L. Xiao, “Solar power generation forecast based on LSTM”, in Proceedings of IEEE International Conference on Software Engineering and Service Science (ICSESS),
2018, pp. 869-872
4

5. Frequency distribution of solar power generation
➢ The maximum values of solar power generation is 8.0 kWh
➢ The minimum values of solar power generation is 0.0 kWh
Solar power generation is strongly biased towards 0.0 kWh
This imbalance dataset
makes it difficult to build an
accurate prediction model
5

6. Average daily solar power generation in each month
Solar power generation is varied over various seasons in one year
This large seasonal
variation of the dataset also
makes it difficult to build
an accurate prediction
model
6
The gap between highest and lowest solar power
generation is very large

7. Data augmentation
➢ The size and variability of training datasets are important factors that affect
the prediction accuracy of deep learning models.
➢ In general, limited size and variability of the training dataset lead to
underfitting or overfitting problems.
○ Data augmentation alleviates those problems by increasing the size and
variability of training datasets with artificially generated samples
➢ Data augmentation for time-series
○ Basic approaches: time-domain and frequency-domain methods
○ Advanced approaches: statistical-based and learning-based methods
7

8. Generative Adversarial Network (GAN)
➢ GAN is one of the most popular learning-based
methods for time-series data augmentation
➢ GAN can prevent some limitations caused by
adversarial learning in the practical application
of the conventional generative models
➢ Two models are trained simultaneously by an
adversarial process.
○ A generator ("the artist") learns to create
images that look real
○ A discriminator ("the art critic") learns to
tell real images apart from fakes.
8
References
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in Proceedings of Advances in Neural Information Processing
Systems (NerIPS), Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Q. Weinberger, Eds., vol. 27. Curran Associates, Inc., 2014.

9. Dataset
➢ Since solar power system does not generate electricity at night, we take only data for
every hour from 6:00 AM to 6:00 PM
○ There are 13 records per day
Dec 6, 2017, 6:00 AM - Jun 17, 2020, 6:00 PM
Past solar power generation data
Solar power system installed at
Shinohashi building of
NTT Facilities (NTT-F)
Weather forecast data [1]
Meso-Scale Model of
Japan Meteorological Agency (JMA)
9
References
[1] K. Saito, T. Fujita, Y. Yamada, J. ichi Ishida, Y. Kumagai, K. Aranami, S. Ohmori, R. Nagasawa, S. Kumagai, C. Muroi, T. Kato, H. Eito, and Y. Yamazaki, “The operational JMA nonhydrostatic
mesoscale model,” Monthly Weather Review, vol. 134, no. 4, pp. 1266–1298, 2006.

10. Correlation of the features in our dataset
The highest correlation coefficient between
Solar radiation & Solar power generation
with a correlation coefficient of 0.79
The lowest correlation coefficient between
Middle cloudage & Solar power generation
with a correlation coefficient of -0.34
Correlation coefficient
It is used to measure the strength
of the relationship between two
variables.
10

11. Our approach
11
Dataset
Training
Dataset
Testing
Dataset
LSTM-based Prediction Model
GAN-based Model
Augmented Dataset
Evaluation
1. Preparation of training and testing datasets
2. Data augmentation using GAN-based model
3. Prediction using LSTM-based model

12. Input and output data range for the prediction model
➢ We propose a model to predict the next 24 hours solar power generation (i.e. 13 hours data
except for the night time) with the past information
t
Current time step
Solar power generation
Weather forecast information
Solar power generation
t + 13
t - 13
➢ The length of input sequence is 13
➢ The length of output sequence is 13
Predict future solar power generation for
next 13 time steps except for the night time.
12

13. Input data for the proposed GAN-based model
The input data format of the proposed GAN-based
model is composed of two parts
13
First part is a matrix of 13 x 12 that matches with the
input data of our proposed LSTM-based model
Second part is a matrix of 13 x 1 that matches with
the output data of our proposed LSTM-based model

14. Architecture of proposed GAN-based model
14
Input layer
Discriminator model
Generator model
Convolutional layer
First
Deconvolutional layer
Second
Deconvolutional layer
First
Convolutional layer
Second
Convolutional layer
Fully connected layer
Input layer

15. Architecture of LSTM-based prediction model
The length of input sequence
The number of selected features
The length of output sequence
The number of hidden units
15
Output layer is fully-connected layer
with ReLU activation function

16. Original Data
Generated Data
Example of generated and original data
The proposed GAN-based model was able to generated data efficiently because the
distribution of the generated data after training process is similar to the original data
16
Generated data at 1st
epoch Generated data at 100th
epoch

17. Prediction accuracy
17
The prediction model using
the original dataset
The prediction model using
the augmented dataset
Since neural networks are initialized with
random weights, we trained the model 100 times
with stratified K-fold cross-validation, and
measured and averaged the RMSE.
0.1497
0.1898
0.0606
0.0802

18. Prediction results
18
The prediction results indicates that the prediction error around noon is greater than
that of early morning and late evening.

19. Conclusion
➢ We studied how to improve the control of stored electricity in batteries to
produce stable electricity by predicting future solar power generation using
the past solar power generation and weather information
➢ The proposed LSTM-based prediction model achieved an RMSE of 0.1898
with stratified k-fold cross-validation
➢ To further enhance the accuracy of our prediction model, we designed a
neural network model based on GAN to augment the training dataset.
○ The proposed generative model is able to increase the number and
variability of the dataset. With the augmented dataset, our prediction
model achieved an RMSE of 0.0802 19

20. Future works
➢ Other data augmentation methods will be investigated to further improve the
prediction accuracy
➢ Data on solar power generation and weather information from other sources
should be used to validate the generality of the proposed prediction method
➢ Significant behaviors and features that impact on the control of the stored
electricity in batteries will be investigated
20

21. Q&A
Thank you for your attention
Email: thonglek.kundjanasith.ti7@is.naist.jp
21