Kun Xiong

Bio: Kun Xiong is an academic researcher. The author has contributed to research in topics: Power system simulation & Dynamic load testing. The author has an hindex of 1, co-authored 1 publications receiving 25 citations.

Dynamic Load Models for Industrial Facilities

Abstract: This paper presents a new method to construct dynamic models for large industrial and commercial facilities commonly connected to power transmission systems. These facilities typically draw large amounts of power and have complex dynamic responses to power system disturbances. Traditional load modeling approaches such as those based on load composition or site measurement are not adequate to produce dynamic models for such facilities. In this paper, a facility template-based load modeling technique along with template scaling/equivalence algorithms is proposed to solve the facility modeling problem. Oil refinery facilities are used as an example to illustrate the proposed modeling technique. The technique requires minimal user input and can be implemented in a database program.

Measurement-Based Load Modeling Using Transfer Functions for Dynamic Simulations

Abstract: Measurement-based load modeling is a promising approach to reliably represent load behavior in dynamic simulations of large power systems. This paper presents a methodology that starts with the acquisition of voltages and currents from power quality monitoring systems and highlights the issues associated with selecting, processing and resampling the data to estimate the relationship between the power deviations as a function of the voltage deviations. The load model mathematical structure chosen is a second-order transfer function, whose parameters are estimated using a genetic algorithm (GA) as the optimization technique that minimizes the error between the real data that are measured and the data that are simulated with the proposed models. Some insights were achieved regarding the appropriate search space choice.

Fast and Reliable Estimation of Composite Load Model Parameters Using Analytical Similarity of Parameter Sensitivity

Abstract: This paper proposes a computationally efficient technique for estimating the composite load model parameters based on analytical similarity of parameter sensitivity When the model parameters are updated in the optimization procedure to best fit the actual load dynamics, ie, measurements, parameters of similar sensitivity representation in the given mathematical model structure are updated in the same manner at every iterative step This research allows for practically reducing the number of load model parameters to be identified in the estimation process and the overall computational cost while preserving the desired complexity and accuracy of the original model This approach consequently facilitates the parameter estimation in the optimization process and helps manage increased number of parameters often criticized for adopting the dynamic composite load model via measurement-based approach Case studies for the real power system demonstrate the computational efficiency and intact accuracy of the proposed method with reference to the existing methods of estimating all the parameters of the given composite load model independently

Universidad de chile facultad de ciencias físicas y matemáticas departamento de ingeniería mecánica

Abstract: Soft Biometrics uses face or iris images to estimate demographic information, such as, gender, ethnicity, age and, emotions. In a biometric recognition system, gender information may lead to searching only half of the database. Most gender classi cation methods reported in the literature use all of the features extracted for classi cation purposes. In image understanding, raw input data often has very high dimensionality and a limited number of samples. In this area, feature selection plays an important role in improving accuracy, e ciency and scalability of the object identi cation process. One approach to feature selection is to compute Mutual Information (MI) by ranking features according to their individual relevance using a lter approach. Feature ranking methods are considered to be fast and e ective, particularly when the number of features is large and the number of available training examples is comparatively small, as in the case of micro-arrays. However, the ranking reduces the selection power of the criterion because it does not compute any form of complementarity among variables, making it ill-suited for feature selection on datasets that have high levels of complementarity. Possible redundancies between variables are not taken into account. Two methods were developed in this thesis, the rst one to select pairs of features and the second one to select groups of features. In the rst method, feature selection methods use pairs of features based on mutual information applied to the gender classi cation problem using frontal facial and, iris images. Fusion is performed for the selected features based on intensity of pixels, shape and texture in several spatial scales. Results were assessed and compared to those previously published on standard databases as the FERET and UND databases for controlled environments and, the Labeled Faces in the Wild (LFW) database for uncontrolled environments. The method was also applied to iris images with near infrared illumination from the University of Notre Dame database, named Gender From Iris (GFI). The second method is based on lters and wrappers ( ltrapper), for groups of feature selection instead of pairs, using mutual information. This method was applied to gender classi cation using face and iris images and to benchmark databases (Sonar, Spambase, Ionosphere and Madelon). The selection of groups of features based on relevant and redundant features was performed using a weight related to neighbor features. The method was applied to face images using the UND database, the LFW and MORPH II databases, and to iris images using the GFI. The results for the rst method on gender classi cation from faces, obtained with the fusion of 18,900 selected features reached a classi cation rate of 99.13% on the FERET database. For the UND database, the best gender classi cation performance was obtained with the fusion of 14,200 selected features reaching a classi cation rate of 94.01%. For the LFW the best performance was obtained with the fusion of 10,400 features from 3 di erent spatial scales obtained a classi cation rate of 98.01%. Depending on the image size, the total number of features was reduced at least 70% on FERET, 73% on the UND database and 90% on LFW database. To our understanding, these results are the best result reported so far for gender classi cations on the databases. In iris, the best results using 10x10 overlapping windows for histogram LBP(8,1) gave the best accuracy, obtaining 91.33%. This level of accuracy exceeds that of any other publication that we are aware of. The results for the second method, the best gender classi cation performance was obtained on the UND database with the fusion of 2,100 selected features and 30 nearest neighbors obtaining a classi cation rate of 98.25%. The best gender classi cation accuracy for the MORPH II database was reached with the best 1,300 selected features and 25 nearest neighbors achieving a classi cation rate of 95.50%. For the LFW dataset the best result reached was 97.00% with 900 features and 25 nearest neighbors also using cross-validation with ve groups for the training set. Our results show that gender classi cation can be improved signi cantly by feature selection using groups of the best features. The best classi cation accuracy for the benchmark databases was Sonar dataset yielding 95.68% using 55 features and 35 nearest neighbors. In feature selection from iris, our approach achieved gender prediction accuracy of 89% based on fusing the best features from the left and the right iris.