考虑低通信成本的张量分解个性化联邦学习超短期电力负荷预测方法

doi:10.19595/j.cnki.1000-6753.tces.251325

摘要
图/表
参考文献
相关文章 (15)

全文: PDF (2541 KB)
输出: BibTeX | EndNote (RIS)

摘要

精确的超短期负荷预测对于电力系统的实时调控与安全稳定运行至关重要。传统的集中式协同建模方法依赖于跨域数据汇集,易引发数据安全问题,而现有联邦学习面临全局模型收敛慢、通信成本高的难题。为此,本文提出一种考虑低通信成本的张量分解个性化联邦学习超短期电力负荷预测方法。首先,使用张量分解技术设计了一种新的张量化本地模型,以降低通信成本;其次,设计了双层损失函数,通过控制个性化模型与张量本地模型之间的距离,实现个性化模型优化与张量本地模型的解耦;最后,根据模型参数张量化特性,设计了一种高效的个性化学习策略和集成张量聚合机制,实现全局模型聚合与更新。通过理论分析与实验验证,表明所提方法在完成快速收敛的同时,有效提升了模型的鲁棒性与泛化能力,增强了预测稳定性,降低了联邦学习模型训练过程中的通信成本。

	服务

	把本文推荐给朋友
	加入我的书架
	加入引用管理器
	E-mail Alert
	RSS
	作者相关文章
	罗光浩
	崔明建
	韩一宁
	黄勇
	张剑

关键词 ：联邦学习, 张量分解, 个性化学习策略, 低通信成本, 负荷预测

Abstract：

With the continuous growth of electricity demand, high-precision ultra-short-term load forecasting plays a vital role in improving grid dispatching efficiency, enhancing operational flexibility, and ensuring the reliability of power supply. However, due to the data island problem, individual power suppliers typically have access to only limited user load data, making it difficult to effectively conduct data-driven load forecasting and thus significantly restricting forecasting accuracy. Although centralized data aggregation by power companies can partially alleviate this issue, the growing emphasis on data security has introduced strict regulatory constraints on centralized data collection. Moreover, the potential risk of data leakage during the transmission of critical equipment data remains a crucial concern. Federated learning (FL) provides an effective solution to the data island problem by enabling cross-domain collaboration through distributed joint modeling under the principle of data locality, thereby avoiding both the risk of data leakage during transmission and the high storage costs associated with centralized data collection. Nevertheless, existing FL frameworks often overlook the communication costs incurred during parameter exchanges between clients and the central server throughout the training process. Furthermore, they fail to fully account for the data heterogeneity among clients during global model aggregation, resulting in slow convergence and limited generalization performance.
To address these challenges, this paper proposes a low-communication-cost tensor decomposition-based personalized federated learning method for ultra-short-term load forecasting. The proposed approach aims to achieve efficient and robust distributed forecasting while ensuring data privacy and security. Firstly, a novel tensorized local model based on tensor decomposition is designed to extract and compress spatiotemporal features, thereby enhancing the representational capacity of model parameters and reducing communication costs during FL training. Secondly, a dual-loss optimization framework is constructed to alternately optimize the personalized model and the tensorized local model, constraining their distance to decouple personalized optimization from global model learning. Finally, leveraging the tensor characteristics of model parameters, an efficient personalized FL training strategy and an integrated tensor aggregation mechanism are developed to achieve efficient model training and aggregation, thereby improving the generalization capability of the global model.
To systematically evaluate the comprehensive performance of the proposed model, a series of multi-dimensional comparative experiments were conducted. Firstly, pFedACT was compared with several state-of-the-art federated learning methods in the field of power load forecasting to assess its capability in capturing the spatiotemporal variability of load data. Secondly, the predictive performance of pFedACT was compared with that of centralized modeling approaches to evaluate its learning effectiveness under privacy-preserving conditions. Thirdly, the model's robustness was verified across four dimensions: data imbalance, temporal granularity, load diversity, and scenario variability. Next, a convergence analysis is conducted from two perspectives, namely convergence stability and convergence efficiency, to evaluate the algorithm's behavior during training. Then, by comparing the performance of the global model and personalized models under heterogeneous client data distributions, the generalization ability of pFedACT was examined. Subsequently, the effectiveness of the proposed method in distributed modeling was then validated by comparing the performance differences between various base models trained in centralized settings and within the pFedACT framework. Furthermore, experiments with varying numbers of clients were conducted to evaluate the model's feasibility and scalability across diverse power system scenarios. Finally, a detailed communication cost analysis was performed to quantify the advantages of pFedACT in terms of communication efficiency.
Extensive simulation results demonstrate that the proposed method not only achieves rapid model convergence but also exhibits excellent robustness and generalization capability across power system datasets under various operating conditions. Meanwhile, it enhances prediction stability and reduces communication costs during the training process. The simulation analysis leads to the following conclusions: (1) The proposed pFedACT employs a dual-loss function to impose regularization between the personalized model and the tensorized model, thereby enhancing the model's adaptability to diverse load scenarios and significantly improving both robustness and convergence speed. (2) The tensor decomposition and integrated tensor aggregation strategies strengthen the model's spatiotemporal representation ability, enabling superior generalization performance under non-independent and identically distributed data conditions. (3) Compared with baseline federated learning methods, pFedACT reduces communication costs by approximately 32.23%, and this advantage becomes increasingly prominent as the number of clients grows or the model complexity increases.

Key words： Federated learning tensor decomposition personalized learning strategy low communication cost load forecasting

收稿日期: 2025-07-26

PACS:

TM743

基金资助:

国家重点研发计划（2023YFB2407500）和国家自然科学基金（52207130）资助项目

通讯作者: 崔明建男,1987年生,教授,博士生导师,研究方向为电力系统运行控制、电力系统优化运行、电力系统数据分析等。E-mail：mjcui@tju.edu.cn

作者简介: 罗光浩男,1995年生,博士研究生,研究方向为人工智能技术在新型电力系统的应用。E-mail：luoguanghao@tju.edu.cn

引用本文:

罗光浩, 崔明建, 韩一宁, 黄勇, 张剑. 考虑低通信成本的张量分解个性化联邦学习超短期电力负荷预测方法[J]. 电工技术学报, 0, (): 20251325-. Luo Guanghao, Cui Mingjian, Han Yining, Huang Yong, Zhang Jian. A Low-Communication-Cost Tensor Decomposition-Based Personalized Federated Learning Method for Ultra-Short-Term Power Load Forecasting. Transactions of China Electrotechnical Society, 0, (): 20251325-.

链接本文:

https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.251325 https://dgjsxb.ces-transaction.com/CN/Y0/V/I/20251325

[1] Akyol H B, Preist C, Schien D.Avoiding overconfidence in predictions of residential energy demand through identification of the persistence forecast effect[J]. IEEE Transactions on Smart Grid, 2022, 14(1): 228-238.
[2] Kim N, Park H, Lee J, et al.Short-term electrical load forecasting with multidimensional feature extraction[J]. IEEE Transactions on Smart Grid, 2022, 13(4): 2999-3013.
[3] 李云松, 张智晟. 考虑综合需求响应的Transformer-图神经网络综合能源系统多元负荷短期预测[J]. 电工技术学报, 2024, 39(19): 6119-6128.
Li Yunsong, Zhang Zhisheng.Transformer based multi load short-term forecasting of integrated energy system considering integrated demand response[J]. Transactions of China Electrotechnical Society, 2024, 39(19): 6119-6128.
[4] de Vilmarest J, Browell J, Fasiolo M, et al. Adaptive probabilistic forecasting of electricity (net-)load[J]. IEEE Transactions on Power Systems, 2023, 39(2): 4154-4163.
[5] 李延珍, 王海鑫, 杨子豪, 等. 基于非侵入式负荷分解的家庭负荷两阶段超短期负荷预测模型[J]. 电工技术学报, 2024, 39(11): 3379-3391.
Li Yanzhen, Wang Haixin, Yang Zihao, et al.Two-stage ultra-short-term load forecasting model of household appliances based on non-intrusive load disaggregation[J]. Transactions of China Electrotechnical Society, 2024, 39(11): 3379-3391.
[6] 赵洪山, 吴雨晨, 温开云, 等. 基于时空注意力机制的台区多用户短期负荷预测[J]. 电工技术学报, 2024, 39(07): 2104-2115.
Zhao Hongshan, Wu Yuchen, Wen Kaiyun, et al.Short-term load forecasting for multiple customers in a station area based on spatial-temporal attention mechanism[J]. Transactions of China Electrotechnical Society, 2024, 39(7): 2104-2115.
[7] 钟吴君, 李培强, 涂春鸣. 基于EEMD-CBAM-BiLSTM的牵引负荷超短期预测[J]. 电工技术学报, 2024, 39(21): 6850-6864.
Zhong Wujun, Li Peiqiang, Tu Chunming.Traction load ultra-short-term forecasting framework based on EEMD-CBAM-BiLSTM[J]. Transactions of China Electrotechnical Society, 2024, 39(21): 6850-6864.
[8] Oreshkin B N, Dudek G, Pełka P, et al.N-BEATS neural network for mid-term electricity load forecasting[J]. Applied Energy, 2021, 293(1): 116918-116930.
[9] 冯昌森, 钱燚飞, 邵亮, 等. 考虑特征缺失的个性化居民短期负荷预测[J]. 电力系统自动化, 2025, 49(16): 75-84.
Feng Changsen, Qian Yifei, Shao Liang, et al.Personalized short-term residential load forecasting considering missing feature[J]. Automation of Electric Power Systems, 2025, 49(16): 75-84.
[10] Husnoo M A, Anwar A, Hosseinzadeh N, et al.A secure federated learning framework for residential short term load forecasting[J]. IEEE Transactions on Smart Grid, 2023, 15(2): 2044-2055.
[11] Zhang Yigong, Cui Qiushi, Shi Lixian, et al.PPenergyNET: Privacy-preserving Multi-energy Load Forecasting in Energy Internet Considering Energy Coupling[J]. IEEE Transactions on Power Systems, 2024, 39(5): 6235-6248.
[12] Wang Renjun, Qiu Haifeng, Gao Hongjun, et al.Adaptive horizontal federated learning-based demand response baseline load estimation[J]. IEEE Transactions on Smart Grid, 2023, 15(2): 1659-1669.
[13] Qin Dalin, Wang Chenxi, Wen Q, et al.Personalized federated darts for electricity load forecasting of individual buildings[J]. IEEE Transactions on Smart Grid, 2023, 14(6): 4888-4901.
[14] Si Caomingzhe, Wang Haijin, Chen Lei, et al.Robust co-modeling for privacy-preserving short-term load forecasting with incongruent load data distributions[J]. IEEE Transactions on Smart Grid, 2024, 15(3): 2985-2999.
[15] He Yu, Luo Fengji, Sun Mingyang, et al.Privacy-preserving and hierarchically federated framework for short-term residential load forecasting[J]. IEEE Transactions on Smart Grid, 2023, 14(6): 4409-4423.
[16] Dinh C T, Tran N H, Nguyen T D.Personalized federated learning with moreau envelopes[J]. Advances in neural information processing systems, 2020, 33: 21394-21405.
[17] Tchaye-Kondi J, Zhai Yanlong, Shen Jun, et al.Adaptive period control for communication efficient and fast convergent federated learning[J]. IEEE Transactions on Mobile Computing, 2024, 23(12): 12572-12586.
[18] Zhang Weishan, Zhou Tao, Lu Qinghua, et al.FedSL: A communication-efficient federated learning with split layer aggregation[J]. IEEE Internet of Things Journal, 2024, 11(9): 15587-15601.
[19] Pei Jiaming, Li Wei, Mumtaz S.From routine to reflection: pruning neural networks in communication-efficient federated learning[J]. IEEE Transactions on Artificial Intelligence, 2024: 1-10.
[20] Deng Yongheng, Lyu Feng, Xia Tengxi, et al.A communication-efficient hierarchical federated learning framework via shaping data distribution at edge[J]. IEEE/ACM Transactions on Networking, 2024, 32(3): 2600-2615.
[21] Bubeck S.Convex optimization: Algorithms and complexity[J]. Foundations and Trends® in Machine Learning, 2015, 8(3-4): 231-357.
[22] Karimireddy S P, Kale S, Mohri M, et al.Scaffold: Stochastic controlled averaging for federated learning[C]. International conference on machine learning. PMLR, 2020: 5132-5143.
[23] Fallah A, Mokhtari A, Ozdaglar A.Personalized federated learning: A meta-learning approach[J]. arXiv preprint arXiv: 2002. 07948, 2020.
[24] Yu Hao, Jin Rong, Yang Sen.On the linear speedup analysis of communication efficient momentum SGD for distributed non-convex optimization[C]. International Conference on Machine Learning. PMLR, 2019: 7184-7193.
[25] Jiang Yuang, Wang Shiqiang, Valls V, et al.Model pruning enables efficient federated learning on edge devices[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 34(12): 10374-10386.
[26] Li Guanzheng, Li Bin, Li Chao, et al.Battery state of charge probabilistic estimation using natural gradient boosting[J]. IEEE Transactions on Industrial Electronics, 2023, 71(9): 10636-10646.
[27] 全睿, 程功, 周宇龙, 等. 基于增强型鲸鱼优化算法CNN-BiGRU-AT模型的燃料电池衰退预测[J]. 电工技术学报, 2025, 40(19): 6342-6358.
Quan Rui, Cheng Gong, Zhou Yulong, et al.Enhanced whale optimization algorithm-based CNN-BiGRU-AT model for aging prediction of fuel cell[J]. Transactions of China Electrotechnical Society, 2025, 40(19): 6342-6358.
[28] Shirin Y.2011-2017-load: Electricity load datasetof areas in Alberta, Canada for 2011-2017[EB/OL]. Hugging Face, 2025-10-16 [2024-10-15]. https://huggingface.co/datasets/ShirinYamani/2011-2017-load.
[29] Bensalah M, Hair A, Rabie R, et al. High-resolution smart meter load dataset collected from multiple cities in Morocco[EB/OL]. UCI Machine Learning Repository, 2025-05-03 [2025-10-15]. https://doi.org/10.24432/C5FC8R.
[30] Muehlenpfordt J. Time series: Load, wind and solar, prices in hourly resolution[EB/OL]. Open Power System Data, 2020-10-06 [2025-10-15]. https://data.open-power-system-data.org/time_series/2020-10-06.