基于动态权重混合专家模型的超短期电力负荷预测

doi:10.11930/j.issn.1004-9649.202506007

中国电力 ›› 2026, Vol. 59 ›› Issue (5): 118-132.DOI: 10.11930/j.issn.1004-9649.202506007

基于动态权重混合专家模型的超短期电力负荷预测

周专¹(), 王杰³(), 边家瑜², 于志勇², 袁铁江³()

1. 国网新疆电力有限公司，乌鲁木齐　830000
2. 国网新疆电力有限公司经济技术研究院，乌鲁木齐　830063
3. 大连理工大学电气工程学院，大连　116081

收稿日期:2025-06-06 修回日期:2025-08-22 发布日期:2026-05-15 出版日期:2026-05-28
作者简介:
周专（1979），男，副高级工程师，从事电力负荷预测研究，E-mail： 345035567@qq.com
王杰（2001），男，通信作者，硕士研究生，从事电力大数据、AI算法、人工智能研究，E-mail：2542935445@qq.com
袁铁江（1976），男，教授，从事电-氢高效转换与互动控制技术、电夹点分析与新型电力系统规划技术、储能与新能源发电并网技术研究，E-mail：ytj1975@dlut.edu.cn
基金资助:
国网新疆经研院科技项目（SGXJ0000FCJS2400181）。

Ultra-short-term power load forecasting based on dynamic weighting mixture of experts

ZHOU Zhuan¹(), WANG Jie³(), BIAN Jiayu², YU Zhiyong², YUAN Tiejiang³()

1. State Grid Xinjiang Electric Power Co., Ltd., Urumqi 830000, China
2. Economic and Technological Research Institute, State Grid Xinjiang Electric Power Co., Ltd., Urumqi 830063, China
3. School of Electrical Engineering, Dalian University of Technology, Dalian 116081, China

Received:2025-06-06 Revised:2025-08-22 Online:2026-05-15 Published:2026-05-28
Supported by:
This work is supported by Science and Technology Project of State Grid Xinjiang Economic Research Institute (No.SGXJ0000FCJS2400181)

摘要/Abstract

摘要：

超短期电力负荷预测是新型电力系统实时调度的关键支撑技术，其精度直接决定新能源消纳能力、机组组合经济性及储能系统充放效率。针对负荷数据的强时序性、气象敏感性、日期敏感性及异常波动挑战，提出动态权重混合专家模型（dynamic weight-mixture of experts，DW-MoE）用于超短期负荷预测。首先，该模型通过双向长短期记忆网络（bi-directional long short-term memory，BiLSTM）捕捉负荷序列的周期性时序特征，借助极端梯度提升树（extreme gradient Boosting，XGBoost）刻画气象因子、日期因子与负荷的非线性关联，利用生成对抗网络（generative adversarial networks，GAN）实现异常负荷模式的精准检测。然后，设计基于滑动窗口误差反馈的动态权重机制，实现多专家输出的自适应融合。最后，引入在线更新机制，基于最新采样数据对模型参数进行增量式优化，提升模型对非平稳负荷波动的动态响应能力。实验结果表明，相较于单一模型及传统混合方法，DW-MoE模型在超短期电力负荷预测精度和收敛速度上均表现出明显优势，尤其在异常负荷场景中预测误差降低显著，验证了模型对突变负荷模式的鲁棒性。

关键词: 超短期负荷预测, 混合专家模型, 动态权重, 在线更新

Abstract:

Ultra-short-term power load forecasting is a key supporting technology for real-time scheduling of new-type power systems, and its accuracy directly determines the consumption capacity of new energy, the economics of units combination, and the charging and discharging efficiency of energy storage system. To address the challenges of load data—including its strong temporal dependency, sensitivity to meteorological conditions, sensitivity to calendar effects, and anomalous fluctuations, a dynamic weighting mixture of experts (DW-MoE) model is proposed for ultra-short-term load forecasting. Firstly, the model captures the periodic temporal characteristics of the load sequences through BiLSTM, characterizes the nonlinear correlation between meteorological factors, date factors and loads using XGBoost, and achieves accurate detection of abnormal load patterns using GAN. Then, a dynamic weighting mechanism based on sliding window error feedback is designed to achieve adaptive fusion of multiple expert outputs; Finally, an online update mechanism is introduced to incrementally optimize the model’s parameters based on the latest sampled data, enhancing the dynamic response capability of the model to non-stationary load fluctuations. The experimental results demonstrate that compared to single models and traditional hybrid methods, the DW-MoE model exhibits significant advantages in both prediction accuracy and convergence speed for ultra-short-term load forecasting, and notably, it achieves a substantial reduction in prediction error under anomalous load scenarios, validating the model's robustness to abrupt load variations.

Key words: ultra-short-term load forecasting, mixture of experts, dynamic weighting, online updates

周专, 王杰, 边家瑜, 于志勇, 袁铁江. 基于动态权重混合专家模型的超短期电力负荷预测[J]. 中国电力, 2026, 59(5): 118-132.

ZHOU Zhuan, WANG Jie, BIAN Jiayu, YU Zhiyong, YUAN Tiejiang. Ultra-short-term power load forecasting based on dynamic weighting mixture of experts[J]. Electric Power, 2026, 59(5): 118-132.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.electricpower.com.cn/CN/10.11930/j.issn.1004-9649.202506007

https://www.electricpower.com.cn/CN/Y2026/V59/I5/118

图/表 11

参考文献 35

1	申洪涛, 李飞, 史轮, 等. 基于气象数据降维与混合深度学习的短期电力负荷预测[J]. 电力建设, 2024, 45 (1): 13- 21.
	SHEN Hongtao, LI Fei, SHI Lun, et al. Short-term power load forecasting based on reduction of meteorological data dimensionality and hybrid deep learning[J]. Electric Power Construction, 2024, 45 (1): 13- 21.
2	于润泽, 窦震海, 张志一, 等. 基于二次分解重构与多任务学习的综合能源系统多元负荷短期预测[J]. 电力建设, 2024, 45 (12): 149- 161.
	YU Runze, DOU Zhenhai, ZHANG Zhiyi, et al. Multi-energy load forecasting of integrated energy system based on secondary decomposition-reconstruction and multi-task learning[J]. Electric Power Construction, 2024, 45 (12): 149- 161.
3	李科, 潘庭龙, 许德智. 基于MSCNN-BiGRU-Attention的短期电力负荷预测[J]. 中国电力, 2025, 58 (6): 10- 18.
	LI Ke, PAN Tinglong, XU Dezhi. Short-term power load forecasting based on MSCNN-BiGRU-Attention[J]. Electric Power, 2025, 58 (6): 10- 18.
4	孟浩, 徐飞, 符帅, 等. 考虑温控型负荷特性影响的集群用户超短期负荷预测方法[J]. 中国电力, 2025, 58 (12): 63- 72, 85.
	MENG Hao, XU Fei, FU Shuai, et al. Ultra-short-term load forecasting method for aggregated users considering the impact of temperature-controlled load characteristics[J]. Electric Power, 2025, 58 (12): 63- 72, 85.
5	吴军英, 路欣, 刘宏, 等. 基于Spearman-GCN-GRU模型的超短期多区域电力负荷预测[J]. 中国电力, 2024, 57 (6): 131- 140.
	WU Junying, LU Xin, LIU Hong, et al. Ultra-short-term multi-region power load forecasting based on spearman-GCN-GRU model[J]. Electric Power, 2024, 57 (6): 131- 140.
6	范士雄, 李东琦, 郭剑波, 等. 基于时变滤波经验模态分解-重构和独立自注意力机制的iTransformer超短期负荷预测方法[J]. 电网技术, 2025, 49 (6): 2436- 2445.
	FAN Shixiong, LI Dongqi, GUO Jianbo, et al. Ultra-short-term load forecasting method based on time-varying filter empirical mode decomposition-reconstruction and iTransformer with stand-alone self-attention mechanism[J]. Power System Technology, 2025, 49 (6): 2436- 2445.
7	王永利, 刘泽强, 董焕然, 等. 基于CEEMDAN-CSO-LSTM-MTL的综合能源系统多元负荷预测[J]. 电力建设, 2025, 46 (1): 72- 85.
	WANG Yongli, LIU Zeqiang, DONG Huanran, et al. Multivariate load forecasting of integrated energy system based on CEEMDAN-CSO-LSTM-MTL[J]. Electric Power Construction, 2025, 46 (1): 72- 85.
8	朱凌建, 荀子涵, 王裕鑫, 等. 基于CNN-Bi LSTM的短期电力负荷预测[J]. 电网技术, 2021, 45 (11): 4532- 4539.
	ZHU Lingjian, XUN Zihan, WANG Yuxin, et al. Short-term power load forecasting based on CNN-BiLSTM[J]. Power System Technology, 2021, 45 (11): 4532- 4539.
9	张晓佳, 王灿, 张佳恒, 等. 基于多能需求响应与改进BiLSTM的综合能源系统负荷预测[J]. 电力建设, 2025, 46 (4): 113- 125.
	ZHANG Xiaojia, WANG Can, ZHANG Jiaheng, et al. Integrated energy system load forecasting based on multi-energy demand response and improved BiLSTM[J]. Electric Power Construction, 2025, 46 (4): 113- 125.
10	徐欣然, 张绍兵, 成苗, 等. 基于多路层次化混合专家模型的轴承故障诊断方法[J]. 计算机应用, 2025, 45 (1): 59- 68.
	XU Xinran, ZHANG Shaobing, CHENG Miao, et al. Bearings fault diagnosis method based on multi-pathed hierarchical mixture-of-experts model[J]. Journal of Computer Applications, 2025, 45 (1): 59- 68.
11	李鹏, 罗湘淳, 孟庆伟, 等. 基于Spearman相关性阈值寻优和VMD-LSTM的用户级综合能源系统超短期负荷预测[J]. 全球能源互联网, 2024, 7 (4): 406- 420.
	LI Peng, LUO Xiangchun, MENG Qingwei, et al. Ultra short-term load forecasting of user level integrated energy system based on spearman threshold optimization and variational mode decomposition and long short-term memory[J]. Journal of Global Energy Interconnection, 2024, 7 (4): 406- 420.
12	茹瑶, 赵永宁, 叶林, 等. 超短期LSTM风电功率预测模型的混合专家模块化代理解释方法[J]. 电力建设, 2024, 45 (11): 114- 124.
	RU Yao, ZHAO Yongning, YE Lin, et al. Modular surrogate interpretation method based on decision tree mixture of experts for ultra-short-term LSTM wind power forecasting model[J]. Electric Power Construction, 2024, 45 (11): 114- 124.
13	杨雪莹, 祁琪, 李启明, 等. 奖励机制与用户意愿结合的高峰期负荷博弈调度策略[J]. 电工技术学报, 2024, 39 (16): 5060- 5074.
	YANG Xueying, QI Qi, LI Qiming, et al. Peak load game scheduling strategy combining reward mechanism and user willingness[J]. Transactions of China Electrotechnical Society, 2024, 39 (16): 5060- 5074.
14	XUE W, HE H, WANG Y B, et al. SAGN: sparse adaptive gated graph neural network with graph regularization for identifying dual-view brain networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2025, 36 (5): 8085- 8099.
15	谈耀荻, 黄艳国, 刘景锋, 等. 基于MMoE-CNN-Informer模型的电力系统多元负荷长短期时间序列预测[J]. 电气工程学报, 2025, 20 (2): 253- 263.
	TAN Yaodi, HUANG Yanguo, LIU Jingfeng, et al. Long and short-term multivariate load forecasting based on MMoE-CNN-Informer model for power systems[J]. Journal of Electrical Engineering, 2025, 20 (2): 253- 263.
16	WANG R F, WU J T, CHENG X, et al. Adaptive expert fusion model for online wind power prediction[J]. Neural Networks, 2025, 184, 107022.
17	程鹏举, 樊艳芳, 侯俊杰, 等. 基于Conformer-MoE的多设备迁移学习非侵入式负荷分解方法[J]. 智慧电力, 2026, 54 (1): 102- 109.
	CHENG Pengju, FAN Yanfang, HOU Junjie, et al. A multi-device transfer learning non-intrusive load disaggregation method based on conformer-MoE[J]. Smart Power, 2026, 54 (1): 102- 109.
18	文越, 邵必林, 曹翔, 等. 基于Transformer-GAN的天然气负荷数据异常检测分析[J]. 集成电路应用, 2025, 42 (1): 58- 61.
	WEN Yue, SHAO Bilin, CAO Xiang, et al. Analysis of anomaly detection of natural gas load data based on Transformer-GAN[J]. Application of IC, 2025, 42 (1): 58- 61.
19	李加文, 孙永辉, 王森, 等. 计及异常场景数据缺失的负荷超短期预测[J]. 电力系统自动化, 2025, 49 (15): 133- 143.
	LI Jiawen, SUN Yonghui, WANG Sen, et al. Ultra-short-term load forecasting considering missing data in anomalous scenarios[J]. Automation of Electric Power Systems, 2025, 49 (15): 133- 143.
20	曾静, 娄冰, 吕娜, 等. 基于多特征和LOF的用户负荷突变检测[J]. 浙江电力, 2023, 42 (2): 90- 97.
	ZENG Jing, LOU Bing, LYU Na, et al. Abrupt user load change detection based on multiple features and LOF algorithm[J]. Zhejiang Electric Power, 2023, 42 (2): 90- 97.
21	YI C C, PENG Y, SU S, et al. Anomaly detection of photovoltaic power generation based on quantile regression recurrent neural network[J]. Electric Power Systems Research, 2025, 238, 111132.
22	刘知秋. 短期电力负荷预测方法研究[D]. 南京: 东南大学, 2023.
	LIU Zhiqiu. Research on short-term power load forecasting method[D]. Nanjing: Southeast University, 2023.
23	CHEN Z Q, YANG Y, JIANG C D, et al. Enhanced industrial heat load forecasting in district networks via a multi-scale fusion ensemble deep learning[J]. Expert Systems with Applications, 2025, 272, 126783.
24	ZHANG K, YANG M W, ZHANG Y M, et al. Error feedback method (EFM) based dimension synthesis optimisation for four-bar linkage mechanism[J]. Applied Soft Computing, 2023, 144, 110424.
25	葛璐璐. 基于滑动窗口的一类非负可变权组合预测方法[D]. 合肥: 安徽大学, 2019.
	GE Lulu. A new non-negative variable weight combination forecasting method based on sliding window[D]. Hefei: Anhui University, 2019.
26	于多, 曹燚, 王海荣, 等. 基于ICEEMDAN-PE和IDBO-Informer组合模型的短期负荷预测[J]. 中国电力, 2025, 58 (6): 19- 32.
	YU Duo, CAO Yi, WANG Hairong, et al. Short-term load forecasting based on a combined ICEEMDAN-PE and IDBO-Informer model[J]. Electric Power, 2025, 58 (6): 19- 32.
27	李笑竹, 王维庆. 基于贝叶斯理论的分布鲁棒优化在储能配置上的应用[J]. 电网技术, 2022, 46 (10): 4001- 4011.
	LI Xiaozhu, WANG Weiqing. Application of distributed robust optimization based on Bayesian theory in allocation of energy storage[J]. Power System Technology, 2022, 46 (10): 4001- 4011.
28	任恒宇, 韩冬, 任曦骏, 等. 基于时变价格弹性矩阵的深谷电价多目标定价策略[J]. 电网技术, 2024, 48 (3): 958- 967.
	REN Hengyu, HAN Dong, REN Xijun, et al. Multi-objective deep valley electricity pricing model based on time-varying price elasticity matrix[J]. Power System Technology, 2024, 48 (3): 958- 967.
29	毛明轩, 冯心营, 陈思宇, 等. 基于贝叶斯优化卷积神经网络的路面光伏阵列最大功率点电压预测方法[J]. 中国电机工程学报, 2024, 44 (2): 620- 630.
	MAO Mingxuan, FENG Xinying, CHEN Siyu, et al. A novel maximum power point voltage forecasting method for pavement photovoltaic array based on Bayesian optimization convolutional neural network[J]. Proceedings of the CSEE, 2024, 44 (2): 620- 630.
30	韩璟琳, 冯喜春, 胡平, 等. 基于Hyperopt-LightGBM的直流配电网短期负荷抗噪声预测[J]. 高电压技术, 2024, 50 (11): 4902- 4911.
	HAN Jinglin, FENG Xichun, HU Ping, et al. Hyperopt-LightGBM-based noise-resistant forecasting of short-term loads in the DC distribution network[J]. High Voltage Engineering, 2024, 50 (11): 4902- 4911.
31	李扬, 马文捷, 卜凡金, 等. 多智能体深度强化学习驱动的跨园区能源交互优化调度[J]. 电力建设, 2024, 45 (5): 59- 70.
	LI Yang, MA Wenjie, BU Fanjin, et al. Deep reinforcement learning-driven cross-community energy interaction optimal scheduling[J]. Electric Power Construction, 2024, 45 (5): 59- 70.
32	李志军, 徐博, 张家安, 等. 基于TD3可变长度时间窗口最优加权的短期负荷预测策略[J]. 电力建设, 2024, 45 (6): 140- 148.
	LI Zhijun, XU Bo, ZHANG Jia’an, et al. Short-term load optimal weighted forecasting strategy based on TD3 variable length time window[J]. Electric Power Construction, 2024, 45 (6): 140- 148.
33	柳华, 熊再豹, 蒋陶宁, 等. 分布式强化学习驱动的微电网群动态能量优化管理策略[J]. 中国电力, 2025, 58 (10): 50- 62.
	LIU Hua, XIONG Zaibao, JIANG Taoning, et al. Distributed reinforcement learning-driven dynamic energy optimization management strategy for microgrid clusters[J]. Electric Power, 2025, 58 (10): 50- 62.
34	陈实, 唐国登, 刘艺洪, 等. 基于条件风险强化学习的梯级水光蓄联合优化调度[J]. 电力建设, 2025, 46 (9): 84- 97.
	CHEN Shi, TANG Guodeng, LIU Yihong, et al. Condition-based risk reinforcement learning for joint optimal scheduling of cascade hydropower and solar-powered reservoirs[J]. Electric Power Construction, 2025, 46 (9): 84- 97.
35	LIU Y X, ZHANG K X, CHEN S S, et al. Interpretable prediction of short-term load curve for extreme scenarios based on mixture of experts model[J]. Electric Power Systems Research, 2025, 247, 111730.

类型	具体特征	维度
时序特征	前 24 h负荷值（15 min间隔，共 96 点）	96
气象特征	温度、水汽压、风速、光照强度	3
时间与日历特征	小时（0~23）、星期几（1~7）、季节（1~4）、节假日（0/1）、工作日（0/1）、	5
异常特征	GAN 判别器输出的异常得分（0~1，值越高表示异常程度越高）	1

类型	具体特征	维度
时序特征	前 24 h负荷值（15 min间隔，共 96 点）	96
气象特征	温度、水汽压、风速、光照强度	3
时间与日历特征	小时（0~23）、星期几（1~7）、季节（1~4）、节假日（0/1）、工作日（0/1）、	5
异常特征	GAN 判别器输出的异常得分（0~1，值越高表示异常程度越高）	1

特征集	RMSE/ MW	MAE/ MW	MAPE/ %	模型训练时间/min
完整特征集	33.15	24.08	2.31	45.2
筛选后特征集	32.72	23.73	2.28	32.6

特征集	RMSE/ MW	MAE/ MW	MAPE/ %	模型训练时间/min
完整特征集	33.15	24.08	2.31	45.2
筛选后特征集	32.72	23.73	2.28	32.6

模型	RMSE/MW	MAE/MW	MAPE/%
LightGBM	57.33	48.55	4.98
GNN	59.22	50.32	5.09
GRU	59.12	50.27	5.08
MDLinear	60.80	51.30	5.11
MDLinear+GRU	58.63	49.22	5.03
Autoformer	56.12	48.12	4.92
FEDformer	55.34	47.21	4.83
iTransformer	53.41	45.20	4.69
MoE	35.69	25.74	2.46
DW-MoE	32.72	23.73	2.28

基于动态权重混合专家模型的超短期电力负荷预测

Ultra-short-term power load forecasting based on dynamic weighting mixture of experts

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 35

相关文章 3

编辑推荐

Metrics

组合	RMSE/MW	MAE/MW	MAPE/%
一	36.84	26.57	2.63
二	35.92	25.89	2.51
三	34.15	24.36	2.35
四	33.87	24.12	2.31
五	38.21	27.45	2.89
六	40.56	29.38	3.12
本文	32.72	23.73	2.28

α 值	RMSE/MW	MAE/MW	MAPE/%
0.10	36.89	27.42	2.61
0.15	34.52	25.17	2.43
0.20	32.72	23.73	2.28
0.25	33.15	24.09	2.32
0.30	35.27	26.31	2.53

S值	A_c/%	异常场景 RMSE/MW	整体RMSE/ MW	计算耗时/ (ms·样本^–1)
0.4	85.7	31.25	34.15	62
0.5	91.2	32.61	32.72	58
0.6	96.3	35.18	33.89	55

模型组件	更新周期	RMSE/ MW	MAE/ MW	MAPE/ %	日均计算耗时/min
BiLSTM	12 h	33.15	24.09	2.32	48.6
	24 h	32.72	23.73	2.28	25.3
	48 h	34.26	25.17	2.41	12.8
XGBoost	5 天	33.02	23.95	2.30	32.5
	7 天	32.72	23.73	2.28	22.1
	10 天	33.89	24.68	2.37	15.7
GAN	15 天	33.27	24.15	2.33	65.4
	30 天	32.72	23.73	2.28	38.2
	60 天	34.59	25.32	2.45	19.3

[1]	崔一宸, 王鹤, 王莉丽, 银涛, 黄山松. 基于改进生成对抗网络场景生成的跨流域水风光互补中长期优化调度方法[J]. 中国电力, 2026, 59(4): 35-46.
[2]	孟浩, 徐飞, 符帅, 孙鹏, 郝玲, 刘博宇, 刘芷维. 考虑温控型负荷特性影响的集群用户超短期负荷预测方法[J]. 中国电力, 2025, 58(12): 63-72, 85.
[3]	罗嘉，陈世和，张曦，吴乐. 电站锅炉典型热工参数软测量研究[J]. 中国电力, 2016, 49(6): 48-52.

模态框（Modal）标题

基于动态权重混合专家模型的超短期电力负荷预测

Ultra-short-term power load forecasting based on dynamic weighting mixture of experts

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 35

相关文章 3

编辑推荐

Metrics