College of Electrical and Power Engineering, Taiyuan University of Technology

【目的】基于集中式数据的深度学习可以有效提高风功率的预测精度,但数据泄露带来的严重后果令各风电场不断重视自身数据的保密,对数据驱动的风功率预测方法造成阻碍。【方法】针对上述问题,提出一种面向数据隐私保护的分布式多风电场短期功率预测方法,利用横向联邦学习(federatedlearning,FL)框架完成对风电场的功率预测任务。首先,多个风电场采用分布训练方式,利用时序模式注意力(temporalpatternattention,TPA)机制和长短期记忆(longshort-termmemory,LSTM)网络组合成TPA-LSTM本地模型完成本地数据训练;再将这些本地模型的参数进行聚合处理,同时引入模型上传权重值,提高拟合效果好的本地模型的贡献率;最后实现全局模型参数的更新。【结果】实验结果证明,提出的方法在保证风电场数据隐私的前提下,获得的全局模型在多场景下具有良好的预测性能和泛化能力。

【Purposes】 Deep learning based on centralized data can effectively improve the pre-diction accuracy of wind power, but the serious consequences of data leakage make wind farms constantly pay attention to the confidentiality of their own data, which hinders data-driven wind power prediction methods. 【Methods】 To solve the above problems, a short-term power predic-tion method was proposed for distributed multi-wind farms oriented to data privacy protection, which uses the Horizontal Federated Learning framework to complete the power prediction task of wind farms. First, wind farms adopt the distributed training method, and use the Temporal Pattern Attention (TPA) mechanism and Long Short-Term Memory (LSTM) network to form a TPA-LSTM local model to complete local data training. Then, the parameters of these local models are aggregated, and the weight values of the models are introduced to improve the contri-bution rate of the local models with good fitting effect. Finally, the update of global model parameters is implemented. 【Findings】 The results of experiments show that the global model ob-tained by this method has good prediction performance and generalization ability in multiple sce-narios under the premise of ensuring the privacy of wind farm data.

新疆维吾尔自治区重大科技专项(2022A01007-2);山西省自然科学青年基金资助项目(202203021212296)

郑杰,牛哲文,韩肖清,等.面向数据隐私保护的分布式多风电场短期功率预测[J].太原理工大学学报,2024,55(1):102-110.

ZHENG Jie,NIU Zhewen,HAN Xiaoqing,et al.Distributed multi-wind farm short-term power prediction for data privacy-preserving[J].Journal of Taiyuan University of Technology,2024,55(1):102-110.