Enhancing the Prediction Accuracy of Solar Power Generation using a Generative Adversarial Network

• Published in: 2021 IEEE Green Energy and Smart Systems Conference (IGESSC) pp. 1 - 6
• Main Authors: Thonglek, Kundjanasith, Ichikawa, Kohei, Takahashi, Keichi, Nakasan, Chawanat, Yuasa, Kazufumi, Babasaki, Tadatoshi, Iida, Hajimu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-11-2021
• Subjects: Adversarial Learning; Conferences; Data Augmentation; Green products; Neural networks; Power supplies; Predictive models; Solar Power Systems; Time-Series Forecasting; Training; Weather forecasting
• ISSN: 2640-0138
• DOI: 10.1109/IGESSC53124.2021.9618702

Solar power is the most widely used green energy. However, using solar power generation as a stable power supply remains challenging since the power output is difficult to predict. Accurate prediction of solar power generation enables efficient control of the amount of stored electricity in batteries to produce a stable supply of electricity. This paper aims to build a highly accurate solar power prediction model. For this purpose, we design a neural network model based on Long Short-Term Memory (LSTM) to predict the future solar power generation using past solar power generation and weather forecasts. Since a large and diverse dataset is required to train an accurate prediction model, we develop a neural network based on Generative Adversarial Network (GAN) to generate artificial datasets from the original training dataset to increase the amount and diversity of the training dataset. Additionally, stratified k-fold cross-validation is used to eliminate learning deviation during training. As a result, the proposed neural network model based on GAN improved the R 2 score of LSTM from 0.750 to 0.805 with stratified k-fold cross-validation.