【摘要】 为有效挖掘光伏发电功率数据中的有效时序信息,进一步提升光伏发电功率预测效果,提出一种基于多因素融合的高效通道注意力机制(ECA)-时间卷积网络(TCN)预测模型。首先,采用最大互信息系数(MIC)提取光伏发电功率相关特征;其次,使用多项式特征衍生方法融合各相关因素特征,衍生高维特征,进行特征组合;然后,将自适应选择一维卷积核大小的ECA模块与可有效捕捉光伏发电功率数据时序性信息的TCN相结合,搭建ECA-TCN预测模型;最后,采用多个模型进行对比实验。实验结果表明:该文提出的特征组合方法可高效的选择光伏发电功率数据特征,提升特征的表现能力。特征组合后的ECA-TCN预测模型的均方根误差(RMSE)为0.0828 kW,相较于LSTM、LSTM-TCN、ECA-LSTM,ECA-TCN的RMSE分别降低了0.29、0.23和0.13个百分点,并具有最优的拟合度(R~2)90.74%。该模型可在保持高拟合度的同时有效提高光伏发电功率预测精度。更多还原

【Abstract】 To effectively extract valuable temporal information from photovoltaic(PV) power data and further improve the accuracy of PV power prediction, we propose an efficient channel attention mechanism(ECA)-temporal convolutional network(TCN) prediction model based on multi-factor fusion. Firstly, the maximum information coefficient(MIC) is utilized to extract relevant features of PV power. Secondly, a polynomial feature derivation method is employed to combine the features of various factors, generating high-dimensional features and facilitating feature combination. Then, the ECA module, which adaptively selects the size of one-dimensional convolutional kernels, is combined with TCN to construct the ECA-TCN prediction model, which effectively captures the temporal characteristics of PV power data. Finally, multiple models are compared through experimental evaluation. The results demonstrate that the proposed feature combination method efficiently selects PV power data features and enhances their discriminative ability. The ECA-TCN prediction model with feature combination achieves a root mean square error(RMSE) of 0.0828 kW. Compared to LSTM, LSTM-TCN, and ECA-LSTM, the RMSE of the ECA-TCN was reduces by 0.29, 0.23, and 0.13 percentage points, respectively, and exhibits the best fitting performance with an R~2 value of 90.74%. This model effectively improves the accuracy of PV power prediction while maintaining a high fitting performance.更多还原