Copyright ©2023Published by IRCS - ITB J. Eng. Technol. Sci. Vol. 55, No. 2, 2023, 109-119 ISSN: 2337-5779 DOI: 10.5614/j.eng.technol.sci.2023.55.2.1 Research Paper Development of Non-Intrusive Load Monitoring of Electricity Load Classification with Low-Frequency Sampling Based on Support Vector Machine Edi Leksono 1, * , Auditio Mandhany 1 , Irsyad Nashirul Haq 1 , Justin Pradipta 1 , Putu Handre Kertha Utama 1 , Reza Fauzi Iskandar 1,2 & Rezky Mahesa Nanda 1 1 Engineering Physics, Institut Teknologi Bandung, Jalan Ganesha No.10, Bandung 40132, Indonesia 2 Engineering Physics, Telkom University, Jalan Telekomunikasi No.1, Bandung 40257, Indonesia Corresponding author: edi@tf.itb.ac.id Abstract Non-intrusive load monitoring (NILM) is a promising approach to provide energy consumption monitoring of electrical appliances and analysis of current and voltage data with less instrumentation. This paper proposes an electrical load classification model using support vector machine (SVM). SVM was chosen to keep the computational cost low and be able to implement an embedded system. The SVM model was utilized to classify the on/off state of air conditioners, light bulbs, other uncategorized electronics, and their combinations. It utilizes low-frequency sampling data captured every minute, or at a 0.0167 Hz rate. Utilization change in active and reactive power was used as a feature in the model training. The optimal kernel for the model was the radial basis function (RBF) kernel with C and gamma values of 88.587 and 2.336 as hyperparameters, producing a highly accurate model. In testing with real-time conditions, the model classified the on/off state of the electrical loads with 0.93 precision, 0.91 recall, and 0.91 f-score. The results of testing proved that the model can be applied in real time with high accuracy and with an acceptable performance in field implementation using an embedded system. Keywords: energy monitoring; load classification; low frequency sampling; non-intrusive load monitoring; support vector machine. Introduction Energy conservation in buildings through energy efficiency optimization is of the utmost importance nowadays, as it is part of climate change mitigation [1,2]. The first step of energy conservation is to monitor energy usage and to profile electrical energy consumption. Thus, the consumer will get information on how much energy they use and why it reaches a particular value. The study by Batra [3] explains that awareness of energy consumption associated with real-time energy observation encourages users to change their energy usage behavior, leading to more sustainable energy consumption. However, the total energy consumption data alone was reported to be ineffective in changing consumers’ energy usage behavior [4]. Energy usage measurement at the appliance level is necessary. Although appliance-level energy measurement can yield very accurate results [5], system deployment is expensive [4]. For this problem, non-intrusive load monitoring (NILM) exists as a solution. In addition to the lower measurement costs, the NILM methodology can be proposed to deduce electrical load information and reduce appliance complexity. NILM is a method for disaggregating total electrical load from one measurement point into individual appliances by using their distinctive characteristics [6]. Load separation can be solved by looking at the load conditions when the utility is turned on/off and operated at varying power states. To classify appliances based on these situations, a certain classification method is needed based on load characteristics. The challenge of NILM lies in determining the characteristics to classify loads. Load characteristics also depend on the frequency of data Journal of Engineering and Technological Sciences