基于CNN-GRU-BiLSTM-AM的超短期光伏功率预测

Ultra-short-Term Photovoltaic Power Prediction Based on CNN-GRU-BiLSTM-AM

摘要:

为提升超短期光伏功率预测精度,提出了一种基于卷积神经网络-门控循环单元-双向长短期记忆神经网络-注意力机制(CNN-GRU-BiLSTM-AM)的组合预测模型。首先,为提升数据质量,对数据的异常值进行处理,采用Spearman相关系数、灰色关联分析提取影响光伏功率的关键特征;其次,为获取光伏功率数据的时空特征,分别利用CNN和GRU提取空间和时间维度上的局部特征,利用BiLSTM捕捉时间序列中的长期依赖关系;再次,为获取关键历史时间点的重要信息,引入AM,构建了CNN-GRU-BiLSTM-AM预测模型。最后,结合公开的光伏功率数据集进行对比实验。结果显示,构建的预测模型决定系数为99.1%,均方根误差为0.032 5,平均绝对误差为0.026 6,表明该方法有效提高了光伏功率的预测精度。

Abstract:

To improve the ultra-short-term prediction accuracy of photovoltaic power, a hybrid prediction model based on CNN-GRU-BiLSTM-AM was proposed. First, to improve data quality, outliers were processed, and the Spearman' s rank correlation coefficient and grey relational analysis were used to extract the key features affecting photovoltaic power. Second, to obtain the spatiotemporal features of photovoltaic power data, the local features in the spatial and temporal dimensions were respectively extracted by CNN and GRU, and the long-term dependence relationships in the time series were captured BiLSTM. Third, to obtain the important information of key historical time points, the Attention Mechanism (AM) was introduced, and the CNN-GRU-BiLSTM-AM prediction model was constructed. Finally, comparative experiments were conducted using a publicly available photovoltaic power dataset. The results show that the prediction model constructed in this study has a coefficient of determination of 99.1%, a root mean square error (RMSE) of 0.032 5, and a mean absolute error (MAE) of 0.026 6, indicating that the method effectively improves the prediction accuracy of photovoltaic power.