.png){kind=icon}
.png){kind=icon}
Return
Probabilistic prediction system based on quantile deep learning model and multi-level information recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126734.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The traditional point prediction methods for photovoltaic power can only provide deterministic information to decision-makers, which fail to effectively quantify the uncertainty associated with output power fluctuations, thereby posing significant risks and challenges to grid integration safety and power scheduling. To address this issue, this paper introduces a multi-layered data processing mechanism, model selection mechanism, and semi- interval optimization mechanism to develop a combined probability forecasting system, which avoids the non- uniqueness of optimal models caused by simultaneously optimizing interval upper and lower bounds. First, in the data processing stage, secondary decomposition extracts information effectively. Second, combining quantile regression with deep learning techniques as alternative models enhances their ability to mine data structures. Finally, improved optimizers perform semi-interval optimization tasks to enhance the global optimization capability and convergence speed. An empirical analysis of data from four photovoltaic power plants in different regions of Australia validates the effectiveness of the combined probability prediction system. The results show that the system achieves the lowest average interval score and quantile loss ranking, with mean ScoreR of 5.875 and standard deviation ScoreStd of 3.140 across all forecasting scenarios compared with other benchmark models. Additionally, the interval coverage and normalized average interval width are also optimal, indicating that the developed system has high reliability and resolution in probability interval forecasting.
Keywords:
Probability forecasting system
Quantile loss
Deep learning models
Semi-interval optimization
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W