摘要
针对目前电力数据维度高、特征复杂、难以进行有效电力用户分析等问题,提出了基于用户电力特征的分割网络模型。首先,基于卷积神经网络(CNN)对电力用户用电特征进行编码,从而分析用户在相邻时段内用电记录的相关性。其次,为了减少信息冗余和高维数据带来的维度爆炸或噪声干扰影响,对电力用户用电记录进行多角度特征提取。接着,基于反卷积网络解码器从提取的特征中重构原始记录,从而可以在无监督的情况下对用户的用电行为进行建模。最后,基于改进粒子群优化(PSO)算法进行超参数优化,从而提高模型训练效率。在仿真阶段,以某电力公司500名用户用电数据为例,对用户用电行为进行分类分析。分析结果表明,经特征提取后,支持向量机(SVM)、变分自编码(VA)、生成对抗网络(GAN)和所提模型轮廓系数分别提升2.58%、4.24%、0.39%和0.86%。经所提改进PSO算法优化后,网络训练性能较传统PSO算法和无优化模型分别提升3.693倍和2.111倍。该模型可用于电力用户分析,为电力企业提高用户服务质量提供一定的借鉴。
The segmentation network model based on user electricity features is proposed to address the current problems such as high dimensionality and complex features of electricity data, which make it difficult to carry out effective analysis of electricity users. Firstly, the electricity consumption features of electricity users are encoded based on convolutional neural network(CNN), to analyze the correlation of users’ electricity consumption records in adjacent time periods. Secondly, to reduce information redundancy and the impact of dimensional explosion or noise interference brought by high-dimensional data, multi-angle features are extracted from the electricity consumption records of electricity users. Then, the original records are reconstructed from the extracted features based on the decoder of the deconvolutional network, so that the electricity consumption behavior of customers can be modeled unsupervised. Finally, hyperparameter optimization is performed based on the improved particle swarm optimization(PSO) algorithm, thus improving the model training efficiency. In the simulation stage, the electricity consumption data of 500 customers of a power company is used as an example to classify and analyze the electricity consumption behavior of customers. The results show that after feature extraction, the support vector machine(SVM), variational self-coding(VA), generative adversarial network(GAN) and the contour coefficient of the proposed model are improved by 2.58%, 4.24%, 0.39% and 0.86%, respectively. After optimization by the proposed improved PSO algorithm, the network training performance is improved by 3.693 times and 2.111 times compared with the traditional PSO algorithm and the no-optimization model, respectively. The model can be used for power customer analysis, which provides a certain reference role for power companies to improve customer service quality.
作者
陈谧
CHEN Mi(Huizhou Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Huizhou 516003,China)
出处
《自动化仪表》
CAS
2022年第9期100-105,共6页
Process Automation Instrumentation
关键词
电力系统
数据分析
用户特征
卷积神经网络
粒子群优化
无监督
特征提取
特征重构
Power system
Data analysis
User characteristics
Convolutional neural network(CNN)
Particle swarm optimization(PSO)
Unsupervised
Feature extraction
Feature reconstruction
