基于动态联邦学习的LA模式下短期负荷预测

Short-Term Load Forecasting in LA Mode Based on Dynamic Federated Learning

摘要: 针对负荷聚合商(LA)模式下数据隐私与样本不足的挑战，本文提出一种动态联邦学习(DFL)与长短期记忆网络(LSTM)融合的短期负荷预测方法。首先，提出一种依托电价合同构建的需求响应机制，以需求响应信号的方式参与负荷预测，构建多维度DR信号模型，引入用户满意度动态约束，优化LA收益目标；其次，建立本地的长短期记忆网络模型。最后设计动态联邦学习(DFL)，通过数据分布感知的权重分配，取代传统平均联邦学习算法，提升模型泛化能力；实验表明，相比传统联邦学习，本文方法在3个场景测试中平均RMSE误差上减少31.2%，MAPE误差降低41.2%，训练轮次减少21.7%，训练时间降低29.6%，在保证各负荷聚合商数据隐私的同时，提升了模型的预测精度，具有更强的实用性。

Abstract: In response to the challenges of data privacy and insufficient samples in the Load Aggregator (LA) model, this paper proposes a short-term load forecasting method that integrates Dynamic Federated Learning (DFL) and Long Short-Term Memory (LSTM) networks. First, a demand response mechanism based on electricity price contracts is proposed, which participates in load forecasting through demand response signals, constructs a multi-dimensional DR signal model, introduces dynamic constraints on user satisfaction, and optimizes the LA revenue target. Second, a local Long Short-Term Memory network model is established. Finally, Dynamic Federated Learning (DFL) is designed to replace the traditional average federated learning algorithm through data distribution-aware weight allocation, enhancing the model’s generalization ability. Experiments show that, compared to traditional federated learning, the proposed method reduces the average RMSE error by 31.2%, the MAPE error by 41.2%, the number of training rounds by 21.7%, and the training time by 29.6% across three scenario tests, while ensuring data privacy for each load aggregator and improving the model’s forecasting accuracy, demonstrating greater practicality.