基于TCN-Wpsformer混合模型的超短期风电功率预测  

Ultra-short-term wind power forecasting based on TCN-Wpsformer hybrid model

作　　者：徐钽 谢开贵[1] 王宇 胡博 邵常政 赵宇生 XU Tan;XIE Kaigui;WANG Yu;HU Bo;SHAO Changzheng;ZHAO Yusheng(State Key Laboratory of Power Transmission Equipment Technology,School of Electrical Engineering,Chonqing University,Chongqing 400044,China)

机构地区：[1]重庆大学电气工程学院输变电装备技术全国重点实验室,重庆400044

出　　处：《电力自动化设备》2024年第8期54-61,共8页Electric Power Automation Equipment

基　　金：国家自然科学基金委员会-国家电网公司智能电网联合基金资助项目(U1966601)。

摘　　要：针对基于梯度下降的递归神经网络难以捕获时间跨度较长的风电功率长期依赖关系的问题,提出一种基于时间卷积网络(TCN)和窗口概率稀疏Transformer(Wpsformer)混合模型的超短期风电功率预测方法。将包含时间季节性特征的时间编码与包含原始数据位置信息的绝对位置编码进行拼接,引入TCN提取时间片段特征,将时间片段特征融入自注意力机制,以时间片段的相关性联系替代时间点的相关性联系。通过Wpsformer模型多步输出超短期风电功率预测值,与原始Transformer模型相比,Wpsformer模型使用窗口概率稀疏自注意力机制,在捕获长期依赖关系的同时筛选出重要程度相对较高的时间片段特征进行计算,提高了预测精度且降低了计算成本。曹店风电场的算例结果表明,所提模型在预测精度方面具有明显优势。消融实验证明了所提模型各模块的必要性。Aiming at the problem that the recurrent neural network based on gradient descent is hard to capture the long-term dependence relationship of wind power with a long time span,an ultra-short-term wind power forecasting method based on the hybrid model of temporal convolutional network(TCN)and window probability sparse Transformer(Wpsformer)is proposed.The time encoding containing the seasonal characteristic and the absolute positional encoding containing the positional information of original data are spliced,the TCN is introduced to extract the time segment features,the time segment features are integrated into the self-attention mechanism,and the relationship between time points is replaced by the relationship between time segments.The ultra-short-term wind power forecasting values are output in multiple stages through the Wpsformer model,compared with the original Transformer model,the Wpsformer model employs the window probability sparse self-attention mechanism to screen out the relative important time segment features for calculation while capture the long-term dependence relationship,which improves the forecasting accuracy and reduces the computational cost.The example results of Caodian wind farm show that the proposed model has obvious advantages in forecasting accuracy.The necessity of each module of the proposed model is verified by the ablation experiment.

关 键 词：超短期风电功率预测 时间卷积网络 窗口概率稀疏Transformer 窗口概率稀疏自注意力机制 

分 类 号：TM614[电气工程—电力系统及自动化]