| 引用本文: | 马丽叶,刘建恒,卢志刚,王海云,袁清芳,杨莉萍.基于深度置信网络的低压台区理论线损计算方法[J].电力自动化设备,2020,40(8): |
| MA Liye,LIU Jianheng,LU Zhigang,WANG Haiyun,YUAN Qingfang,YANG Liping.Theoretical line loss calculation method of low voltage transform district based on deep belief network[J].Electric Power Automation Equipment,2020,40(8): |
|
| 摘要: |
| 针对因线路分布复杂、终端数目庞大等带来的低压台区理论线损计算困难的问题，提出一种基于深度置信网络(DBN)的低压台区理论线损计算新方法。在训练过程中，先利用贪婪算法对DBN模型中的神经网络层进行逐层无监督的预训练，再对该模型进行有监督的全局微调。为了提高计算精度，采用自适应时刻估计(Adam)优化器。以某地区实测2140个台区数据为样本进行仿真计算，结果表明，相较于浅层神经网络，DBN线损计算模型具有更好的泛化能力以及准确性和快速性，且Adam优化器在线损计算中相较于均方根反向传播(RMSProp)和随机梯度下降(SGD)算法具有优越性。 |
| 关键词: 低压台区 理论线损 深度置信网络 贪婪逐层训练法 自适应时刻估计 |
| DOI：10.16081/j.epae.202007036 |
| 分类号:TM74 |
| 基金项目:国家自然科学基金资助项目(61873225) |
|
| Theoretical line loss calculation method of low voltage transform district based on deep belief network |
|
MA Liye1, LIU Jianheng1, LU Zhigang1, WANG Haiyun2, YUAN Qingfang2, YANG Liping2
|
|
1.Key Laboratory of Power Electronics for Energy Conservation and Motor Drive of Hebei Province, Yanshan University, Qinhuangdao 066004, China;2.Electric Power Research Institute of State Grid Beijing Electric Power Company, Beijing 100075, China
|
| Abstract: |
| Aiming at the problem that the theoretical line loss calculation of low voltage transform district is difficult brought by complex distribution of lines and huge number of terminals, a novel calculation method of theoretical line loss based on DBN(Deep Belief Network) is proposed for low voltage transform district. In the training process, the greedy algorithm is used to carry out unsupervised layer-by-layer pre-training of the neural network layer in DBN model first, and then supervised global fine-tuning training of the model is implemented. Adam(Adaptive moment estimation) optimizer is adopted to improve the calculation accuracy. The measured 2140 data of the transform district in a certain area are taken as the samples for simulation and calculation, and results show that the DBN line loss calculation model has better generalization ability, accuracy and rapidity compared with the shallow neural network, and Adam optimizer is superior to RMSProp(Root Mean Square Prop) and SGD(Stochastic Gradient Descent) in line loss calculation. |
| Key words: low voltage transform district theoretical line loss DBN greedy layer training algorithm Adam |
