| 引用本文: | 武新章,赵子巍,代伟,谢代钰,郭苏杭,王泽宇,张冬冬.基于改进的Transformer神经网络辅助的两阶段机组组合决策方法[J].电力自动化设备,2023,43(3): |
| WU Xinzhang,ZHAO Ziwei,DAI Wei,XIE Daiyu,GUO Suhang,WANG Zeyu,ZHANG Dongdong.Two-stage unit commitment decision-making method based on auxiliary of improved Transformer neural network[J].Electric Power Automation Equipment,2023,43(3): |
|
| 摘要: |
| 为了解决大规模电力系统机组组合的“维数灾”问题，提出基于Transformer神经网络的两阶段机组组合决策方法，该方法兼顾求解精度与速度。在第一阶段，考虑机组组合时段耦合的特性，提出基于多头注意力机制的特征向量构建方法，进而基于Transformer神经网络的全局视野与并行化优势，提出一种改进的 Transformer神经网络来预辨识机组启停值。在第二阶段，基于预辨识的机组状态设计置信度阈值，并将机组启停判定可信度定义为启停可信与启停不可信状态，对于启停可信机组的状态进行直接确定，对于启停不可信机组的状态，通过机组组合物理模型进行求解来保证求解的可行性。IEEE 30节点和IEEE 2 383节点系统的仿真结果验证了所提方法的有效性。 |
| 关键词: Transformer神经网络 深度学习 机组组合 数据驱动 特征构造 |
| DOI：10.16081/j.epae.202209014 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金资助项目(52107082)；广西自然科学基金资助项目(2021GXNSFBA220032) |
|
| Two-stage unit commitment decision-making method based on auxiliary of improved Transformer neural network |
|
WU Xinzhang1, ZHAO Ziwei1, DAI Wei1, XIE Daiyu2, GUO Suhang1, WANG Zeyu1, ZHANG Dongdong1
|
|
1.College of Electrical Engineering, Guangxi University, Nanning 530004, China;2.Power Dispatching Control Center of Guangxi Power Grid Co.,Ltd.,Nanning 530023, China
|
| Abstract: |
| In order to solve the “curse of dimensionality” problem of unit commitment in large-scale power system, a two-stage unit commitment decision-making method based on Transformer neural network is proposed, which considers both the solving accuracy and speed. In the first stage, considering the coupling characteristic of unit commitment periods, a feature vector construction method based on multi-Attention mechanism is proposed, further an improved Transformer neural network is proposed to predetermine the unit start-stop values based on the advantages of global view and parallelization of Transformer neural network. In the second stage, the credibility threshold is designed based on the predetermined unit states, and the unit start-stop determination credibility is defined as start-stop credible and start-stop incredible states, the state of start-stop credible unit is determined directly, while the state of start-stop incredible state unit is solved by the unit commitment physical model to ensure the solving feasibility. The simulative results of IEEE 30-bus and IEEE 2 383-bus systems verify the effectiveness of the proposed method. |
| Key words: Transformer neural network deep learning unit commitment data driven feature construction |
