| 引用本文: | 王璐,肖先勇,汪颖,刘阳.基于深度神经网络的电压暂降经济损失评估模型[J].电力自动化设备,2020,40(6): |
| WANG Lu,XIAO Xianyong,WANG Ying,LIU Yang.DNN-based estimation model of economic loss caused by voltage sag[J].Electric Power Automation Equipment,2020,40(6): |
|
| 摘要: |
| 为进一步简化电压暂降经济损失评估流程、提高经济损失预测的适用性和准确度，提出了一种基于深度神经网络(DNN)的电压暂降经济损失评估模型。首先分析了影响电压暂降经济损失的特征因子，分别从电压暂降故障信息、工业过程信息、敏感设备信息和用户基本信息中提取19维特征向量作为DNN预测模型的输入向量，将经济损失结果作为输出，并基于Tensorflow深度学习架构对DNN预测模型进行训练。在此基础上，提出2种数据增强的策略，有效解决了电压暂降样本数据少的窘境，并通过构建4种DNN架构，对比了不同随机失活概率、神经元数量、架构深度对经济损失预测准确度的影响。训练后的DNN模型可以准确提取特征，快速实现收敛并对经济损失进行合理预测。最后，基于我国某大型电子工业企业的电压暂降实际采样数据，对DNN模型进行了训练和性能评估，结果表明了所提方法的有效性。 |
| 关键词: 电压暂降 经济损失 深度神经网络 随机失活层 数据增强 |
| DOI：10.16081/j.epae.202003016 |
| 分类号:TM761 |
| 基金项目:国家自然科学基金资助项目(51807126) |
|
| DNN-based estimation model of economic loss caused by voltage sag |
|
WANG Lu, XIAO Xianyong, WANG Ying, LIU Yang
|
|
College of Electrical Engineering, Sichuan University, Chengdu 610065, China
|
| Abstract: |
| In order to further simplify the process of economic loss assessment of voltage sag and improve the applicability and accuracy of economic loss prediction, an estimation model based on DNN(Deep Neural Network) for economic loss caused by voltage sag is proposed. The characteristic factors affecting the economic loss of voltage sag are analyzed. The 19-dimensional feature vectors are extracted from the sag information, industrial process information, sensitive equipment information and users’ basic information as input vectors of DNN prediction model, and the economic loss results are taken as output. The DNN prediction model is trained based on Tensorflow deep learning framework. On this basis, two data augmentation strategies are proposed to effectively solve the dilemma of few sample data of voltage sag. By constructing four DNN architectures, the effects of different random inactivation probability, number of neurons and depth of architecture on the accuracy of economic loss prediction are compared. As a result, the trained DNN model can extract features accurately, converge quickly and predict economic losses reasonably. Finally, the DNN model is trained and evaluated based on the actual voltage sag sampling data of a large electronic industry enterprise in China, which shows the effectiveness of the proposed method. |
| Key words: voltage sag economic loss deep neural network dropout layer data augmentation |
