Page 122 - 《中国电力》2026年第5期

第 59 卷 第 5 期 Vol. 59, No. 5
2026 年 5 月 ELECTRIC POWER May 2026

引用格式：周专, 王杰, 边家瑜, 等. 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.
Citation: ZHOU Zhuan, WANG Jie, BIAN Jiayu, et al. Ultra-short-term power load forecasting based on dynamic weighting mixture of experts[J].
Electric Power, 2026, 59(5): 118−132.

基于动态权重混合专家模型的超短期

电力负荷预测

周专 ，王杰 ，边家瑜 ，于志勇 ，袁铁江 3
2
3
1
2
(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)
Ultra-short-term power load forecasting based on dynamic weighting
mixture of experts

2
1
3
2
ZHOU Zhuan , WANG Jie , BIAN Jiayu , YU Zhiyong , YUAN Tiejiang 3
(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)

Abstract: Ultra-short-term power load forecasting is a key compared to single models and traditional hybrid methods, the
supporting technology for real-time scheduling of new-type DW-MoE model exhibits significant advantages in both
power systems, and its accuracy directly determines the prediction accuracy and convergence speed for ultra-short-term
consumption capacity of new energy, the economics of units load forecasting, and notably, it achieves a substantial reduction
combination, and the charging and discharging efficiency of in prediction error under anomalous load scenarios, validating
energy storage system. To address the challenges of load the model's robustness to abrupt load variations.
data—including its strong temporal dependency, sensitivity to This work is supported by Science and Technology Project of
meteorological conditions, sensitivity to calendar effects, and State Grid Xinjiang Economic Research Institute (No.SGXJ
anomalous fluctuations, a dynamic weighting mixture of 0000FCJS2400181)
experts (DW-MoE) model is proposed for ultra-short-term load Keywords: ultra-short-term load forecasting; mixture of
forecasting. Firstly, the model captures the periodic temporal experts; dynamic weighting; online updates
characteristics of the load sequences through BiLSTM,
characterizes the nonlinear correlation between meteorological 摘 要：超短期电力负荷预测是新型电力系统实时调度
factors, date factors and loads using XGBoost, and achieves 的关键支撑技术，其精度直接决定新能源消纳能力、机
accurate detection of abnormal load patterns using GAN. Then, 组组合经济性及储能系统充放效率。针对负荷数据的强
a dynamic weighting mechanism based on sliding window error 时序性、气象敏感性、日期敏感性及异常波动挑战，提
feedback is designed to achieve adaptive fusion of multiple 出动态权重混合专家模型（dynamic weight-mixture of experts，
expert outputs; Finally, an online update mechanism is DW-MoE）用于超短期负荷预测。首先，该模型通过双向长
introduced to incrementally optimize the model’s parameters 短期记忆网络（bi-directional long short-term memory，BiLSTM）
based on the latest sampled data, enhancing the dynamic 捕捉负荷序列的周期性时序特征，借助极端梯度提升树
response capability of the model to non-stationary load （extreme gradient Boosting，XGBoost）刻画气象因子、日期
fluctuations. The experimental results demonstrate that 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
收稿日期：2025−06−06； 修回日期：2025−08−22。
多专家输出的自适应融合。最后，引入在线更新机制，基
基金项目：国网新疆经研院科技项目（SGXJ0000FCJS 于最新采样数据对模型参数进行增量式优化，提升模型
2400181）。 对非平稳负荷波动的动态响应能力。实验结果表明，相

118