At present, electric power companys’ judgment on abnomal line loss is that the line loss is abnormal when the line loss rate exceeds a certain threshold. Yet the judgement is one-sidedness and limited. To effectively identify the problem of line loss, based on the study of clustering algorithm and the characteristics of the line loss rate data, an improved k-means clustering algorithm for anomal line loss discrimination is proposed. The method firstly carries out a k-means clustering on the line loss rate of the low voltage substation area to be classified into three classes, then judges whether to carry out secondary classification according to the quantity of various data, and finally judges whether a line loss abnormality exists in the low voltage substation area according to factors such as the size of average line loss rate, and the distance of the clustering center. By analyzing the time dispersion of the class of data with high line loss rate of the clustering results. the degree of abnormality of the line loss can be obtained. Experimental results show that this method has a certain practical application effect and can improve the accuracy of abnormal line loss judgement.

Due to the rapid development of the power industry and the reform of the power market, power companies put higher requirement on power economy. Network losses reduction, especially in the distribution network, has become an urgent problem that need to be solved. Using machine learning as an entry point and through data-driven method, based on different impact on line loss result of different flow, line loss model is built with artificial neural network (ANN), which realizes the theoretical calculation of line loss and electricity theft judgment. Through the relevant cases, the accuracy and reliability of line loss prediction and electricity theft analysis based on ANN model are proved.

The clustering technique in data mining has been widely applied for load curves clustering. Load curves clustering helps refining common and different characteristics among loads, which has important application values for the practicality of load model. On the other hand, it helps analyzing load patterns, guiding planning and real-time dispatching of power systems. In this paper, an adaptive k-means++algorithm is proposed, which synthesizes results of different cluster numbers to verify the similarity of the samples in the dataset, and adopts an iterative graph-partitioning method to identify the optimal cluster number. The improved algorithm avoids excessive deviation of single clustering result caused by inappropriate cluster number of daily load curves, which could improve the accuracy of load curves classification. Numerical experiments verify the feasibility and effectiveness of the proposed algorithm, and show that the accuracy of algorithm is high and robustness is good when solving the best cluster number.

With the development of deep learning technology, the non-intrusive load identification algorithm based on deep learning has become a research hotspot. In this paper, bilateral long-term and short-term memory network (Bi-LSTM) is used for load identification for the first time. A non-intrusive load identification algorithm based on Bi-LSTM is proposed. By locating the moment of load event, the steady state information (active power, reactive power and 15th odd-even current harmonics) of the load operation state are combined as the input of Bi-LSTM which is trained. And precision, recall, accuracy and F1 values are used as evaluation indicators. The results show that the method can identify low-power and multi-state appliances, and Bi-LSTM has stronger recognition ability than long-term and short-term memory network (LSTM) and recurrent neural network (RNN). For the case where there are multiple electrical appliances running in a time period, a method based on the starting and ending feature matching of the load running state is proposed, and the feasibility of the algorithm is proved through experiments.

With the wide application of infrared thermal imaging detection technology in substation inspection robot and transmission line unmanned aerial vehicle (UAV) detection platform, a large number of infrared images with abnormal hot spots on transmission and transformation equipment need to be manually classified and diagnosed periodically, therefore, intelligent diagnosis of these images by intelligent algorithms are urgently needed. At present, classical machine learning algorithms are difficult to effectively identify the abnormal hot spots on infrared image of transmission and transformation equipment. Based on the artificial intelligence deep learning theory, this paper uses Faster RCNN algorithm based on region recommendation network in depth learning algorithm system to detect, identify and locate the heating fault on such infrared images. Based on the image database of heating fault of power transmission and transformation equipment collected by infrared thermal imager, the data set is manually labeled with bounding frames, and the network shared parameters are constructed through alternate training, and the infrared intelligent detection model of abnormal heating of power transmission and transformation equipment is constructed. The method described here provides a new idea for infrared thermography intelligent detection of transmission and transformation equipment.

Due to the larger ground resistance, extinguishing and restrike arc, when high impedance faults of distribution network happen, the high impedance fault (HIF) of distribution network has strong randomness and small amplitude fault current, and traditional over-current protection device cannot detect and work. Artificial intelligence (AI) technology improves the sensitivity and accuracy of HIF detection. In this paper, firstly, the construction method of HIF detection database is introduced. Then AI based HIF detection is analyzed and discussed from signal acquisition, feature extraction and classifier selection. Finally, the key problems in the application of AI to HIF detection are summarized, which provide a solution for the subsequent related researches.

In this paper, a prediction approach for abnormal behavior of substation staff is proposed by combining global and local information using generative adversarial networks based on video scenes. In substations, this method can be used to issue timely warnings to operating and maintenance personnel who may trigger dangerous actions in the process of operation, which provides an important guarantee for the safety of the workers. The prediction task of human behavior aims to predict future action video frames based on given of human action video frames. Considering that human action videos contain both background scenes which are relatively time-invariant among frames, and alseo human actions information which are time-varying components in videos, a global generative adversarial network is used to model the time-invariant background and coarse human profiles; Then a local generative adversarial network is utilized to furtherly refine the time-varying human body details in the video. Compared with existing methods which only obtain pixel-level action predictions, these experiments indicate that our global and local combined approach can capture both spatial appearances and temporal dynamics of videos simultaneously.

AMS (advanced metering system) communicates with vast smart meters via GPRS to collect metering data and send control message, is exposed to the security risk of cyber attack and sensitive data leakage. The application of cyber security situational awareness technology in AMS is helpful to deeply and comprehensively analyze and evaluate the security status of the system. Based on the analysis of the security threats of AMS, a multi-source situational awareness, analysis and evaluation scheme is proposed, which integrates network security, system security, data security and other factors. This paper analyses and designs a unified model of multi-source detection log information collection and formatting, eliminating the problems of relatively isolated detection data and insufficient evidence after the existing system is attacked. Based on the multi-source data fusion and the application of fuzzy reasoning technology, the correlation and risk of network attacks are deduced, which enables the system to detect complex network attacks and evaluate the correlation attacks.

LCL output filter has been widely used in the grid-connected inverter owing to its good high-frequency harmonic attenuation performance. However, the impedance characteristic of the LCL filter has a resonance peak, which tends to cause unstable control of the grid-connected inverter. In order to effectively suppress the resonance, the inverter-side current feedback is introduced as the active damping and the mathematical model of the LCL-filtered inverter under feedback control is obtained. Since the filter parameters and control parameters are coupled to each other, the design for optimizing the filter performance while meeting the demand of control performance becomes particularly complicated. This paper puts forward an overall design method of optimizing the filter parameters and control parameters based on the particle swarm optimization algorithm. The output characteristics of the main harmonics are taken as the optimization goal in the optimal design, with consideration of the filtering performance and control performance. The simulation results verify the feasibility and effectiveness of the proposed method.

With the rapid development of distribution network, a large number of automation equipment are continuously connected to and used in it, which brings great challenges to the reliable operation of distribution network. In this paper, two kinds of machine learning algorithms are proposed to predict the reliability of distribution network according to the dynamic law and reliability index characteristics of distribution network. Through data transformation of distribution network data, the prediction interval is changed from [0,1] to [0,+∞), and the transformed data are input to the predict model again. The results of distribution network example validation show that the prediction effect of machine learning is improved obviously after data transformation, and the artificial neural network optimized by dropout technology has the best prediction performance. The model presented in this paper can accurately predict the reliability of distribution network and provide reference for construction investment and optimal operation of distribution network.