对用户开展精细化用电行为画像及分类,是电力企业精准掌握用户用电规律、提升服务水平和市场竞争力的关键因素之一。针对当前电力用户分类研究中用户用电行为画像结果片面、集成学习负荷分类研究中的基分类器冗余问题及负荷类别不平衡问题,提出一种基于用电行为数字特征画像的电力用户两阶段分类算法。第一阶段,提出一种结合谱聚类和集成强基分类器的用户日负荷曲线分类算法:首先,针对集成学习基分类器学习能力弱的不足,提出一种基于改进长短期记忆网络(long short-term memory,LSTM)的强基分类器;其次,针对基分类器冗余问题,提出一种基于最小正则化代理经验风险的优化选择集成策略;然后,提出一种基于密度的高斯过采样方法处理类别不平衡。第二阶段,基于负荷曲线分类结果,构建以日负荷模式发生概率为数字特征的用户用电行为画像,采用谱聚类算法对用户画像实施分类。最后,通过实测用户负荷数据验证了所提方法的有效性。

Fine power consumption behavior portrait and classification has been one of the key factors for power enterprises to accurately grasp the electricity consumption law of power consumers, improve service level and market competitiveness. To solve the issues of fragmentary portraits of electricity consumption behavior in current power user classification research, base classifier redundancy and class imbalance in ensemble learning load classification, a two-stage power consumer classification method based on digital feature portraits of power consumption behavior is proposed. In the first stage, a classification method for power user daily load curves is proposed combining spectral clustering and integrated strong base classifier. Firstly, a strong base classifier is developed on the basis of LSTM network to improve the weak learning capability of base classifier in ensemble learning. Secondly, an Optimal Selection Ensemble (OSE) strategy based on minimum regularized surrogate empirical risk is proposed to solve the problem of base classifier redundancy. Thirdly, a Density Based Gaussian Synthetic (DBGS) minority over-sampling technique is proposed for class imbalance. In the second stage, the power consumption behavior portraits with daily load-pattern occurrence probability as the digital features are constructed, and the portraits are classified by spectral clustering. Finally, the effectiveness of the proposed method is verified by the measured user load data.

全面两孩政策的实施将对未来中国人口与老龄化变动趋势产生深远的影响，文章在 充分考虑生育政策调整影响的基础上，依托 2015 年 1%人口抽样调查数据对中国 2015 ～2100 年的人 口规模和结构的变动趋势进行预测分析。结果显示: 中国总人口规模将于 2029 年左右迎来峰值，此 后将步入负增长时期，整个人口的年龄结构将不断老化; 劳动年龄人口规模将处于持续缩减之中，劳 动年龄人口比例在21 世纪前半叶将快速下降、后半叶则将在波动中缓慢下降; 老年人口规模将不断 攀升，于 2053 年左右达到峰值，此后将开始逐年下降，老龄化程度在 21 世纪前半叶将快速提升，但到 了 21 世纪后半叶，老龄化进程将逐渐放缓。

非侵入负荷监测是全面感知负荷数据及能效优化的有效途径。当前非侵入式负荷监测算法的主要观测对象是具有调控潜力的负荷,但对于其中功率较小、负荷曲线相似的电器辨识准确率还不够理想,算法对先验数据的依赖程度较高。基于此,提出一种基于多特征联合稀疏表达的SOM-K-means非侵入式负荷辨识算法,该算法利用负荷特征训练得出最优字典,结合最优字典与多特征联合稀疏表示构建目标函数,求解多特征联合稀疏矩阵,克服了单类负荷特征限制识别负荷种类的问题;将多特征联合稀疏矩阵作为输入,结合自组织(self-organizing map, SOM)神经网络优化的K-means算法与平均绝对误差值进行快速辨识。最后,利用PLAID数据集进行了实验验证,结果表明,所提算法仅需迭代120次辨识准确率即可达到90%,提高了算法收敛速度,证明了该方法能够准确高效地实现负荷辨识。

Nonintrusive load monitoring is an effective method for comprehensively perceiving load data and optimizing energy efficiency. At present, the main observation object of nonintrusive load monitoring algorithms is the load with a regulation potential; however, the identification accuracy is poor for electrical appliances with small power and similar load curves. Moreover, the algorithm is highly dependent on prior data. Therefore, an SOM-K-means non-intrusive load identification algorithm based on multi-feature joint sparse expression is proposed in this study. The algorithm uses load features to train the optimal dictionary. The objective function is constructed by combining the optimal dictionary and multi-feature joint sparse representation, and the multi-feature joint sparse matrix is solved, which overcomes the problem of identifying load types limited by single-type load characteristics. Considering the multi-feature joint sparse matrix as the input, combined with the K-means algorithm optimized by a self-organizing map (SOM) neural network and the mean absolute error, the load was quickly identified. Finally, experimental verification using the PLAID dataset shows that the identification accuracy of the proposed algorithm can reach 90% with only 120 iterations, improving the convergence speed of the algorithm and proving that the method can realize load identification accurately and efficiently.

为了降低电力公司的运营成本,针对非技术性损失(non-technical loss, NTL),提出一种基于置信规则推理(belief rule-based,BRB)和长短记忆网络模型(long short-term memory,LSTM)的用户窃电行为诊断方法。该方法首先从用电量大数据中提取电量波动系数和用电量曲线的毛刺宽度两种用电异常特征,制定BRB异常特征输入前置属性转换,并通过证据推理(evidential reasoning,ER)方法输出最终的置信度,建立适用于NTL异常检测的置信规则库,从而自动获取具有高鲁棒性的有标签正负样本训练数据集。接着,以此为基础,提出构建一种多LSTM网络检测模型,实现对异常用电特征的有效提取与检测。实验结果表明,与现有的主流网络故障检测模型相比,所提方法能够更准确地从电力大数据中诊断出用户的异常用电行为。

In order to reduce the operating cost of power companies, for non-technical loss (NTL), this paper proposes a belief rule-based (BRB) and long short-term memory (LSTM) based method for diagnosing the user’s electricity theft behavior. This method firstly extracts the power fluctuation coefficient and the burr width of the power consumption curve from the electric power dataset, and then the input pre-attribute conversion method for BRB anomaly feature is established. The final confidence is output by evidential reasoning (ER) method, and the confidence-rule base for NTL anomaly detection is established, thus, the labeled positive and negative training data sets with high robustness can be obtained automatically. Then, on this basis, a multi-LSTM network detection model is proposed to realize the effective extraction and detection of abnormal electrical features. The experimental results show that, compared with existing mainstream fault detection networks, the proposed method can better accurately diagnose the abnormal electrical behavior of users from the power big data.

现有用户用电模式提取技术主要基于负荷数据时域特征提取，无法准确分辨时域上欧式距离接近但频域上波动特性差异较大的负荷数据，且对类别不平衡负荷数据的分类准确率较低。为解决上述问题，文章首先通过基于样本支持向量的过采样方法（support vector machines-synthetic minority over-sampling technique, SVM-SMOTE）对存在类别不平衡问题的负荷数据进行处理；然后，通过极大重叠离散小波变换（maximal overlap discrete wavelet transform, MODWT）对负荷数据进行分解，并将分解后的尺度系数和细节系数组成频域的特征矩阵；最后将频域特征矩阵输入深度长短期记忆（long short-term memory, LSTM）神经网络进行负荷分类并通过求取各个类别质心来获取典型用电模式。实验结果表明，该方法具有良好的类别不平衡数据处理能力和负荷分类效果。

The current user mode extraction technology is mainly based on the time-domain characteristics of load data. It cannot accurately distinguish load data with close Euclidean distance in time domain and different fluctuation characteristics in frequency domain, and the classification accuracy of imbalanced data is low. This paper proposes a user mode extraction model to solve the above problems. The model firstly processes the data set with class imbalance by SVM-SMOTE over-sampling method, secondly obtains the scale and wavelet coefficients which is used to form the characteristic matrix in frequency domain through overlapping discrete wavelet transform, finally inputs the characteristic matrix into the deep LSTM network for classification and obtains the typical user mode by calculating the centroid of each category. The experimental results show that the method can effectively process imbalanced data and increase classification accuracy.