Research On Data Mining Of Power System Operation Information

The amount of data sampled, accumulated and to be analyzed in power system is expanding drastically, along with the development of information technology and the progress of power grid digitalization. However, limited valuable knowledge can be extracted using inquiry and statistics operation in database or traditional data anlysis method. Therefore it is highly significant to discuss how to obtain hidden pattern and rules from large-scale, multi-dimensional data, in order to provide decision support to power system decision maker.Data mining technique advanced significaly in the past few years with the contradiction between "mass data" and "limited knowledge", and drew attention from diverse researchers. This technique combines various theories such as intellectual algorithms, pattern recognition and mathematics statistics, and is able to discover the rules or knowledge embeded in tremendous data, prior unknown but valuable in decision making. Power system should operate reliably, economically and securely. For these three basic requirements, this thesis conducted intensive study in data mining of power system operation information (voltage and active power flow). The research work accomplished is shown as follows:(1) Phase Space Reconstruction (PSR) method was introduced to power quality research field. The types of concerned disturbances included voltage sags, voltage swells, voltage interruptions, impulsive transients, harmonics and flickers. Based on this method, trajectories of disturbance signals were constructed in phase plane and converted to binary images after normalization and coding process. Four indices of trajectory image which are Maximum Adjacent Distance, Carrier Component Similarity, Overlay Area and Mean Amplitude were presented. Analysis and simulation results showed that by these four indices, features of different disturbances can be extracted effectively. The proposed approach provides a new idea for power quality analysis. (2) The first data mining task in this thesis was conducted for the high-quality requirement of power grid operation. The extracted features using PSR method were utilized as inputs to the support vector machines (SVM) classifier to realize the automatic classification of power disturbances. The types of disturbances discussed included a combination of short-term disturbances (voltage sags, swells) and long-term disturbances (flickers, harmonics), as well as their homologous single ones. Numerical results showed that the method proposed can effectively classify different disturbance patterns and provide basis for power quality mitigation measures. In this research, Comparison studies based on Wavelet Transform (WT) and Artificial Neural Network (ANN) were also reported, to show the advantage of the classification system based on PSR and SVM algorithm.(3) Load profile clustering of power customers is the basis to construct proper tariff system and apply load manage measures. The second data mining task was to study on load profile clustering of low-voltage terminal customers based on Self-Organizing Map (SOM) neural networks. First, three types of vectors which are power curve, time sharing power and power spectrum were defined, and then used as inputs of SOM neural networks to visualize clustering. Two indices, namely relative quantization error and topology error were introduced to evaluate clustering quality. We chose SOM output layer of best performance and allocate load profiles with k-means method. The 131 profiles concerned in this paper were allocated into eight clusters according to Davies index, and each group of profiles is described. Finally, the ability of SOM neural network to identify new customers was examined. The result showed that the method proposed is effective and reliable, and able to supply useful knowledge to enhance power grid operation economy aspect.(4) The third research job for power system operation information was establishing decision tree rules for power grid security assessment. The studied models included Western Systems Coordinating Council (WSCC) three plants nine nodes simplified model and a realistic power grid model in Zhejiang Province. This case study had two goals. First, using knowledge database that covers all possible pre-fault operating conditions, decision rules in the form of hierarchical trees were developed for on-line assessment. Second, Phasor Measurement Units (PMU) were taken into consideration to better decision tree's performance. The results demonstrated that the proposed machine learning scheme was able to identify crucial security indicators and gave reliable security predictions to 96.5% accuracy. Furthermore, voltage phase angle difference that obtained by PMU was proved helpful in improving DT's identification accuracy.The research contents are presented according to steps of data preparation-data mining-interpretation and evaluation. This work expands the application of data mining techniques in power system. The analysis results show that the methods proposed are reliable, comprehensive and practical, and can provide directive significance and application value to enhance the quality, economy and security level of large scale power grid.