Open AccessA new lower and upper bound estimation model using gradient descend training method for wind speed interval prediction
Author(s) -
Liu Fangjie,
Li Chaoshun,
Xu Yanhe,
Tang Geng,
Xie Yuying
Publication year - 2021
Publication title -
wind energy
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.743
H-Index - 92
eISSN - 1099-1824
pISSN - 1095-4244
DOI - 10.1002/we.2574
Subject(s) - interval (graph theory) , wind power , wind speed , prediction interval , gradient descent , upper and lower bounds , function (biology) , renewable energy , computer science , energy (signal processing) , power (physics) , mathematical optimization , control theory (sociology) , algorithm , mathematics , engineering , statistics , artificial neural network , meteorology , artificial intelligence , machine learning , mathematical analysis , physics , electrical engineering , control (management) , combinatorics , quantum mechanics , evolutionary biology , biology
As a clean and renewable energy source, wind energy has achieved remarkable growth around the world. Wind power/speed interval prediction has become an indispensable area of focus regarding the efficient dispatch of wind energy. As an important interval prediction method, the traditional lower and upper bound estimation (LUBE) has been a prevalent approach and a fundamental branch of energy prediction. However, the traditional LUBE model suffers from a low training efficiency owing to a lack of the gradient descent (GD) training mechanism. In this study, an improved LUBE model was designed using a novel training scheme based on the GD method for better efficiency and greater prediction performance. Initially, the new objective functions, which are continuous and differential, meeting the requirements of the GD method, were designed to obtain the best prediction interval (PI) quality with a narrower PI width and greater coverage probability. Then, different loss function forms have been proposed and compared, with the new Huber loss function having been confirmed to be more effective than other traditional loss functions. Finally, the new LUBE model with an objective part and adapting to the GD training method was constructed. Both traditional and improved LUBE models with different loss functions were compared experimentally, and the results indicate that the improved LUBE model with a Huber loss function significantly reduces the training time and improves the quality of the PI.