基于自适应参数聚合的联邦双分支图神经网络及其在用电量需求预测中的应用

Federated 2-Branch Graph Neural Network Based on Adaptive Parameter Aggregation and Its Application in Electricity Demand Forecasting

摘要: 用电量需求预测是智能电网领域的重要研究方向，需要对用电量数据的时序特征和峰值变化进行精确建模。然而，不同地区的用电量数据特性及其需求存在差异，且部分地区的用电情况属于敏感行业数据，需要进行隐私保护。联邦学习为用电量需求预测任务提供了一种兼顾隐私保护的可行方案，但传统的联邦学习方法在该任务中面临两大挑战：一是当本地客户端的用电量数据不足时，其本地模型容易过拟合到含噪声的峰值变化；二是不同客户端的数据特征可能存在显著差异，部分客户端的用电量模式以周期性为主，另一部分则具有明显的峰值变化，导致不同客户端的本地模型参数不一致。相应地，在服务器端聚合这些差异较大的模型会导致全局模型在部分区域的用电数据上出现过平滑现象，降低其对峰值变化的拟合能力。为克服这些挑战，本文提出了一种基于自适应参数聚合的联邦双分支图神经网络(Federated Dual-branch Graph Neural Network with adaptive parameter aggregation, FDGN)用于用电量需求预测。对于局部模型的构建，FDGN首先采用一种混合特征表示，捕捉用电量数据的时序特性和数值波动特性。鉴于时序模式和峰值变化具有不同的数据特征且对预测均至关重要，FDGN采用双分支图对其分别建模：首先设计了用电量时序图及其对应的多尺度图注意力网络以提取用电量数据的时序模式特征，然后设计数值相关图及其对应的高斯自编码器，将峰值变化建模为参数化高斯分布中的动态协方差特征，从而缓解过拟合问题。最后，融合这两部分特征以生成最终预测结果。对于全局模型的构建，FDGN通过在服务器端引入基于相似度的自适应动态参数聚合方法，缓解全局模型的过平滑问题。实验结果表明，在用电量需求预测中，FDGN优于常用的联邦学习方法。

Abstract: Electricity demand forecasting stands as a critical research area within smart grids, requiring accurate modeling of the temporal characteristics and peak variations inherent in consumption patterns. However, significant disparities exist in electricity data characteristics and demand patterns across different regions, with consumption data from certain areas constituting sensitive industrial data necessitating robust privacy protection. Federated learning offers a viable privacy-preserving solution for electricity demand forecasting tasks. Nevertheless, traditional federated learning approaches encounter two primary challenges in this context: 1) local models at clients with insufficient electricity data are prone to overfitting on noisy peak variations; and 2) substantial heterogeneity in data features exists among clients, leading to parameter inconsistency across local models. Consequently, aggregating these divergent models at the server results in an over-smoothed global model that compromises its ability to capture peak fluctuations effectively. To overcome these challenges, this paper proposes a Federated Dual-branch Graph Neural Network with adaptive parameter aggregation (FDGN) tailored for electricity demand forecasting. Within the local model framework, FDGN first constructs a hybrid feature representation capturing both the temporal evolution and numerical volatility of electricity consumption data. Recognizing that temporal patterns and peak variations possess distinct characteristics yet are equally crucial for prediction, FDGN employs a dual-branch graph for parallel modeling: it designs an electricity time-series graph coupled with a corresponding multi-scale attention network to extract temporal patterns, and constructs a numerical structure graph paired with a Gaussian autoencoder to represent peak variations as dynamic covariance features within a parameterized Gaussian distribution, thereby mitigating overfitting. These temporal features and covariance features are ultimately fused to generate the final prediction. For global model construction, FDGN addresses the over-smoothing issue during aggregation by introducing a similarity-based adaptive dynamic parameter aggregation mechanism at the server. Experimental results verify that FDGN outperforms conventional federated learning methods in electricity demand forecasting.