Time series prediction for output of multi-region solar power plants

• Published in: Applied energy 2020-01, Vol.257, p.114001, Article 114001
• Main Authors: Zheng, Jianqin, Zhang, Haoran, Dai, Yuanhao, Wang, Bohong, Zheng, Taicheng, Liao, Qi, Liang, Yongtu, Zhang, Fengwei, Song, Xuan
• Format: Article
• Language: English
• Subjects: Long short-term memory; Multi-region; Particle swarm optimization algorithm; Sensitivity analysis; Solar power output prediction; Time series
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2019.114001

•Long short-term memory is used to predict the multi-region solar power output.•Particle swarm optimization is used to optimize the long short-term memory model.•Sensitivity analysis is done to verify the stability of the method.•Different long short-term memory structures are compared.•A real case is studied to illustrate the effectiveness of the proposed method. Solar energy, as a renewable and clean energy source, has developed rapidly and has attracted considerable attention. The integration of solar energy into a power grid requires precise prediction of the power output of solar plants. Accurate solar power output prediction can promote power dispatch, maintaining the normal operation of power systems. However, research on multi-region solar power is still rare. In this study, long short-term memory and a particle swarm optimization algorithm contribute to solar power prediction considering time series. In order to improve the prediction accuracy, particle swarm optimization is used to optimize the parameters of the long short-term memory model. In addition, different long short-term memory structures are illustrated to determine the final prediction model with sensitivity analysis. Experiments are carried out to verify the effectiveness of the proposed method. The mean absolute error and root mean square error of the proposed method is the smallest among the prediction methods in four cases containing different seasons. In terms of prediction accuracy, results indicate that the proposed prediction model outperforms basic long short-term memory, artificial neural network, and extreme gradient boosting.