| 引用本文: | 庞昊,高金峰,杜耀恒.基于多神经网络融合的短期负荷预测方法[J].电力自动化设备,2020,40(6): |
| PANG Hao,GAO Jinfeng,DU Yaoheng.Short-term load forecasting method based on fusion of multiple neural networks[J].Electric Power Automation Equipment,2020,40(6): |
|
| 摘要: |
| 为了利用不同深度神经网络的优势，提高深度学习算法对短期负荷的预测能力，提出一种基于多神经网络融合的短期负荷预测方法。以电力系统历史有功负荷、季节、日期类型和气象数据为输入特征，并行架构的深度神经网络和注意力机制网络为核心网络；以并行架构中的卷积神经网络通道提取静态特征，门控循环单元网络通道挖掘动态时序特征，采用注意力机制网络融合提取的特征并动态调整网络对不同特征的依赖程度；使用Maxout网络增强网络整体的非线性映射能力，通过全连接网络输出预测结果。与支持向量机、长短期记忆网络的算例结果对比表明，所提方法具有更高的预测平稳性和准确性。 |
| 关键词: 短期负荷预测 多神经网络融合 门控循环单元网络 卷积神经网络 注意力机制网络 Maxout网络 |
| DOI：10.16081/j.epae.202005021 |
| 分类号:TM732 |
| 基金项目: |
|
| Short-term load forecasting method based on fusion of multiple neural networks |
|
PANG Hao1, GAO Jinfeng1, DU Yaoheng2
|
|
1.Industrial Technology Research Institute, Zhengzhou University, Zhengzhou 450001, China;2.Yantai Power Supply Company of State Grid Shandong Electric Power Company, Yantai 264000, China
|
| Abstract: |
| In order to make use of the advantages of different deep neural networks and improve the ability of deep learning algorithm for short-term load forecasting, a short-term load forecasting method based on multiple neural networks fusion is proposed. The historical active power load, season, date type and weather data of power system are taken as input characteristics, while the deep neural network and attention mechanism network of parallel architecture are taken as core network. The static features are extracted by the convolutional neural network channel in the parallel architecture, the dynamic time series features are mined by the gated recurrent unit network channel, and the attention mechanism network is adopted to fuse the extracted features and dynamically adjust the dependent degree of network on different features. Maxout network is used to enhance the non-linear mapping ability of the whole network, and the forecasting results are output through the fully connected network. Compared with the results of support vector machine and long-and short-term memory network, the proposed method has higher forecasting stability and accuracy. |
| Key words: short-term load forecasting fusion of multiple neural networks gated recurrent unit network convolutional neural network attention mechanism network Maxout network |
