针对风电数据的波动性及数据质量问题,提出了一种基于数据清洗与串联门控循环单元(Gated recurrent unit,GRU)及注意力机制(Attention,Att)的风电功率预测模型,并通过联邦学习实现区域模型参数传递.引入 Moreau 信封作为联邦学习的正则化损失函数,提出了个性化联邦学习框架(Personalized fderated larning with Moreau evelopes,PFedMe).兼顾数据安全和预测性能的双重需求,为新建电厂在数据缺失条件下预测风电功率提供了方案.最后,采用百度动态空间风电功率预测挑战赛数据集进行实验验证,结果表明:所提方案在满足隐私安全约束下实现了高效模型训练与预测性能提升.
A Study of Wind Power Prediction Based on Data Cleaning and Personalized Federated Learning
In response to the volatility and quality issues of wind power data,a wind power prediction model based on data cleaning and series of gated recurrent unit(GRU)and attention mechanisms(Att)is proposed,and the parameters of regional model are transferred through federated learning.The Moreau envelope as a regularized loss function for federal learning is introduced,and the personalized federated learning with Moreau envelopes(PFedMe)is proposed.Taking into account the dual needs of data security and predictive performance,and providing a solution for new power plants to predict wind power.Finally,the dataset of Baidu dynamic space wind power prediction challenge is used for experimental verification.The results show that the proposed scheme achieves efficient model training and performance improvement while satisfying privacy and security constraints.
wind power predictiongated recurrent unitfederated learningMoreau envelopes
万方数据
维普
