Open AccessUltra‐short‐term irradiance forecasting model based on ground‐based cloud image and deep learning algorithm
Author(s) -
Zhen Zhao,
Zhang Xuemin,
Mei Shengwei,
Chang Xiqiang,
Chai Hua,
Yin Rui,
Wang Fei
Publication year - 2021
Publication title -
iet renewable power generation
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.005
H-Index - 76
eISSN - 1752-1424
pISSN - 1752-1416
DOI - 10.1049/rpg2.12280
Subject(s) - computer science , irradiance , benchmark (surveying) , cloud computing , feature (linguistics) , algorithm , term (time) , solar irradiance , artificial neural network , artificial intelligence , convolutional neural network , remote sensing , meteorology , geography , linguistics , philosophy , physics , geodesy , quantum mechanics , operating system
Solar irradiance is the main factor affecting the output of a photovoltaic (PV) power station, which is chiefly determined by the cloud distribution over the power station. For ultra‐short‐term, especially the intro‐hour time scale irradiance forecasting, ground‐based cloud image is considered as a very necessary data as Global Horizontal Irradiance (GHI). However, the information content in the image is much higher than that of GHI record, and there is even a difference in magnitude between them. Therefore, how to effectively extract the key features in the cloud images and fuse them with GHI record data is the decisive factor affecting the performance of the forecasting model. Here, a novel convolutional auto‐encoder based cloud distribution feature (CDF) extraction method is first proposed. Then for feature fusion part, an LSTM‐FUSION irradiance forecasting model is established based on long short‐term memory (LSTM) neural network and feature fusion by time steps considering the one‐to‐one correlation between CDFs and GHI. Finally, a novel determination method of input time step length based on attention distribution analysis is also proposed. Simulation results show that the proposed LSTM‐FUSION model is overall superior to the benchmark models.