基于多级特征提取的BiLSTM短期光伏出力预测

SHORT-TERM PHOTOVOLTAIC OUTPUT PREDICTION BASED ON MULTI-LEVEL FEATURE EXTRACTION USING BILSTM

传统光伏预测模型易受气象数据波动的影响,且对气象特征不敏感。由此,提出基于多级特征提取的BiLSTM短期光伏出力预测方法,用于预测不同天气类型下的光伏出力。首先,选取与光伏出力相关性较高的气象因素作为输入特征;使用模糊C均值(FCM)聚类方法,对样本进行灵活划分,通过计算Xie-Beni指标以确定最佳聚类数,将历史数据集聚类为晴天、少云天、晴转多云、阴雨天和恶劣天气;其次,构建CNN-CBAM-TCN多级特征提取器(MFE):利用卷积神经网络(CNN)进行初步的特征提取,结合卷积注意力块(CBAM)抑制非重要特征,之后,利用时间卷积网络(TCN)进一步捕捉日内光伏出力的时序特征;最后,借助双向长短期记忆网络(BiLSTM)进行光伏出力预测。在实例分析中,验证了使用Xie-Beni指标确定最佳聚类数的有效性,证明了该模型较其他预测模型在复杂天气类型下具有更高预测精度。

Traditional photovoltaic(PV)prediction models are highly susceptible to fluctuations in meteorological data and exhibit low sensitivity to meteorological features.To address this,we propose a short-term PV output prediction method based on multi-level feature extraction using bi-directional long short-term memory(BiLSTM),aimed at predicting PV output under various weather conditions.Firstly,meteorological factors with high correlation to PV output are selected as input features.The fuzzy C-means(FCM)clustering method is used for flexible sample division,and the Xie-Beni index is calculated to determine the optimal number of clusters,categorizing historical data into sunny,partly cloudy,cloudy,rainy,and severe weather conditions.Next,a multi-level feature extractor(MFE)comprising CNN-CBAM-TCN is constructed:convolutional neural networks(CNN)are employed for initial feature extraction,convolutional block attention module(CBAM)is used to suppress non-essential features,and temporal convolutional networks(TCN)are utilized to capture the temporal characteristics of intra-day PV output.Finally,BiLSTM is used for PV output prediction.Case studies validate the effectiveness of using the Xie-Beni index to determine the optimal number of clusters and demonstrate that this model achieves higher prediction accuracy compared to other prediction models under complex weather conditions.