基于匿名性差分隐私联邦学习的负荷预测模型训练方法

A Training Method for Load Forecasting Models Based on Anonymity Differential Privacy Federated Learning

随着电力市场改革的不断进行,负荷预测成为了至关重要的一环。电力系统中的边缘设备为了确保用电数据的隐私保护,倾向于将数据保存在本地,这样会在一定程度上阻碍多设备合作训练负荷预测模型,而联邦学习技术可以在数据保存在本地的条件下完成负荷预测模型训练。但目前的研究表明训练的中央服务器可以推测客户端数据集导致隐私泄露。为此,文章提出了客户端-第三方服务器-服务器具有匿名性隐私保护的联邦学习框架,实现了匿名性传输模型参数的同时确保通信效率,并提出了改进的本地化差分隐私联邦学习(local differential privacy federated learning,LDP-FL)算法,确保客户端的隐私安全,同时利用余弦函数筛选参与聚合的梯度,排除恶意客户端。实验结果表明,本方法具有较高的准确率和收敛速率,对于提升电力负荷预测技术的深化发展具有显著的促进作用。

With the continuous reform of the electricity market,load forecasting has become a crucial aspect.Edge devices in the power system tend to keep the data locally to ensure the privacy protection of electricity consumption data,which will hinder the multi-device cooperation in training load forecasting models to some extent,while the federated learning technique can complete the load forecasting model training under the condition that the data is kept locally.However,current research suggests that a centralized server for training can speculate client datasets leading to privacy breaches.For this reason,this paper proposes a client-third-party server-server federated learning framework with anonymity privacy protection,which achieves anonymity in transmitting model parameters while ensuring communication efficiency,and an improved localized differential privacy federated learning(LDP-FL)algorithm to ensure the privacy security of the clients,while utilizing the cosine function to filter the gradients involved in aggregation to exclude malicious clients.The experimental results show that the proposed method has high accuracy and convergence rate,and has a significant role in promoting the deepening of power load forecasting technology.