自样本特征构造的1DCNN-BiLSTM网侧光伏功率预测

1DCNN-BiLSTM Method of Grid-side Photovoltaic Power Prediction with Self-sampled Feature Construction

为解决电网难以获取NWP数据和无法建立光伏功率预测模型的问题,提出一种自样本特征构造的一维卷积双向长短期记忆神经网络光伏发电功率预测方法。通过K均值聚类和功率骤减事件检测的特征工程获取细粒度的天气状态标签,实现基于自身样本的特征构造,以解决样本特征缺少问题;采用卷积和长短期记忆网络结合的模型结构,解决局部特征提取和长期依赖的问题。算例验证结果表明,所提方法改善整体的预测性能,降低多特征数据存在的数据匮乏和数据稳定性风险,为模型输入特征较少的网侧光伏功率短期预测提供一种有效途径。

Aimed at the problem that it is difficult for the power grid to obtain numerical weather prediction(NWP)data and establish a photovoltaic(PV)power prediction model,a one dimensional convolutional neural network bidirectional long short-term memory(1DCNN-BiLSTM)prediction method for PV power generation based on self-sampled feature construction is proposed in this paper.The fine-grained weather status tags are obtained through the feature engineering of K-means clustering and power slump event detection,thus realizing the self-sampled feature construction and solving the problem of lack of features.Moreover,CNN and LSTM networks are combined to solve the problems of local feature extraction and long-term dependency.The validation result of an example indicates that the proposed method improves the overall prediction performance,reduces the data scarcity and stability risks of multi-feature data,and provides an effective approach for short-term prediction of grid-side PV power with few model input features.