基于组合数据清洗与改进注意力机制的光伏发电功率预测

Photovoltaic power forecasting based on combined data cleaning and improved attention mechanism

光伏发电的波动性、间歇性和随机性给电力系统带来巨大挑战,准确预测光伏发电功率是系统安全稳定运行的重大需求,但是目前面临历史数据清洗和预测特征提取两大难题。利用大数据和人工智能技术,提出历史数据组合清洗和改进注意力机制预测方法。首先,研究对预测准确性影响最大的离散型、堆积型异常数据特征,提出一种孤立森林和密度聚类结合的数据分类清洗方法;其次,研究不同类型天气关键数据特征,在现有人工智能算法基础上,建立改进CNN-LSTM-attention光伏发电功率预测模型,提升对不同类型天气的适应能力;最后,针对某实际光伏电站开展实证研究,结果表明,所提出的样本数据组合清洗技术有效提升了异常数据清洗能力,提出的改进注意力机制预测方法预测准确性高,对不同类型天气适应能力强,具有明显的理论价值和工程意义。

The volatility,intermittency,and randomness of photovoltaic power generation pose significant challenges to the power system.Accurately forecasting photovoltaic power generation is a major requirement for the safe and stable operation of the system,However,it is currently faced with two major problems:historical data cleaning and forecasting feature extraction.This paper proposes a method based on combined data cleaning and improved attention mechanism using big data and artificial intelligence technology.Firstly,the study focuses on discrete and cumulative abnormal data features that have the greatest impact on prediction accuracy,proposing a data stratification classification cleaning method combining isolation forests and DBSCAN.Secondly,investigates different types of key weather data features and,based on existing artificial intelligence algorithms,develops an CNN-LSTM-improved attention photovoltaic power forecasting model,enhances adaptability to different weather conditions.Finally,empirical research is conducted on a specific photovoltaic power station,showing that the proposed sample data combination cleaning technology effectively enhances the ability to clean abnormal data,and the improved attention mechanism forecasting method has high accuracy and strong adaptability to different types of weather,with significant theoretical value and engineering significance.