Pradya Panyainkaew creator Weerakorn Ongsakul Chairperson Singh, Jai Govind Examination Committee Than Lin Examination Committee Warodom Khamphanchai Examination committee PEA Scholarship donor Asian Institute of Technology Education Cooperation Project Scholarship donor Royal Thai Government Fellowship Scholarship donor text series technical report th Pathum Thani, Thailand Asian Institute of Technology 2018 continuing eng 59 leaves : ill. (some col.) + 1 online resource In this thesis, support vector machine (SVM) is proposed to identify irregular power consumption which can lead to non-technical loss (NTL) in power distribution system, NTL include faulty metering, equipment failure and electrical fraud. The classifier uses customers{u2019} historical power consumption information in 2016 to investigate suspicious instances which cause irregular power consumption behavior. SVM and ANN require training set data of power consumption for classifier development and test set data for evaluation performance. Training set data contains314 irregular power usage instances and 500 regular power consumption instances. Test set data consistsof100 irregular power consumption instances and the other500 regular ones. Moreover, these information are divided into two scenarios, 249 weekdays and 117 weekend/holidays in 2016, respectively. Every instances in both scenarios are represented by individual average power consumption over 96 fifteen-minute interval a day. To represent consumption characteristic as a probability distribution function, Gaussian mixture distribution which is a feature extraction method, is derived from average power consumption. To cluster various power consumption patterns with the same characteristic, k-means clustering method is applied to both the average power consumption over 96 intervals and Gaussian mixture distribution of combined training and test set data. Using training set of data, SVM classifier is developed by creating a linear hyperplane to separate irregular and regular power consumption instances from each other with maximum margin between both regular and irregular power consumption instances boundary. Subsequently, the classifier with a higher than 85% detection rate of each cluster is used to identify irregular power consumption instances in the same cluster of testing set data based on the area under ROC curve (AUC) and accuracy/detection rate criteria. For feature extraction comparison, SVM with Gaussian mixture distribution provides a higher AUC and accuracy than SVM with the average power consumption for both weekday and holidays. To compare with ANN, SVM with Gaussian mixture distribution render a higher accuracy of 92-95% than ANN with both Gaussian mixture distribution (88-92%) and average power consumption (87-91%) for both weekday and weekend/holidays scenarios. SVM with Gaussian mixture distribution is potentially viable to irregular power consumption identification for distribution utilities. A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Energy Thesis (M.Eng.) - Asian Institute of Technology, 2018 Support vector machines Neural networks (Computer science) Energy consumption Asian Institute of Technology. http://203.159.5.9/ait-thesis/detail.php?q=B02373 http://203.159.5.9/ait-thesis/detail.php?q=B02373 200605 20260818144440.0