REAL TIME ELECTRIC LOAD DISAGGREGATION USING DECISION TREE LEARNING IN AN INTERNET OF THINGS(IOT)

• Main Author: Hafizhankarim Wachid, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55016
• Institution: Institut Teknologi Bandung

Energy is one of the human life important support today, especially electrical energy. Almost all equipment in everyday life uses electric energy, just like at home. The International Energy Agency (IEA) states that electricity consumption in 2017 reached 21,372 TWh. That is 2.6% higher than consumption in 2016. With the increasing use of electronic goods, the need for electrical energy, especially on a home scale, is increasing rapidly. This need can be addressed in two ways, namely building new power plants or reducing electricity consumption itself. Because the construction of new power plants requires a lot of funds and has the potential to damage the environment, saving electricity is the most relevant solution today. In this research, a house energy consumption monitoring system will be built using energy disaggregation from data taken from Non-Intrusive Load Monitoring (NILM). This system is integrated in Internet of Things (IoT) technology so that it can be displayed with a real time user interface. In this study, two disaggregation methods were used, namely combinatorial optimization (minimization) and Classification and Regression Tree (CART). The method used will be validated using a simulation of five lamps operating in the lab. This validation system is built from an energy meter, MQTT communication, a server as a data processing place, and a user interface as a data display. In the validation process, it was found that the CART method has an error of 21 out of 1849 data and produces an accuracy of 98.86%. The method that produces the best accuracy is the CART method. So the method that will be used next for data visualization using Grafana is the CART method. After validation, the method will be used for the AMPDs dataset containing the devices in a house. In the implementation of the system, the system is added by anticipating errors in the field. This anticipation system contains the addition of new devices and changes in device power data. So that the system will recognize whether there is a new device or not. In this system, it has succeeded in recognizing devices with considerable power (50W and 110W). However, on devices with low power (5W) the system cannot detect it. In addition, the change in power was successfully carried out by supervising it when taking data. Each device is turned on individually for power data. After that, the previously validated system is implemented into household data (AMPDs). In household data, there are several appliances that have many states. So that the minimization method is less suitable for this data. Furthermore, the Support Vector Machine (SVM) and minimation method is used as a comparison with the CART method. The CART method accuracy is 96.24%, SVM method accuracy is 73.26%, and minimation metofe acuracy is 47,42%. From these data it can be concluded that the CART method has the best performance to be implemented in household data (AMPDs). After that, visualization of real-time data simulation is done with the result data from the CART method. In the user interface that is made there are four parts, namely total power and state, total energy and the percentage of device energy, a graph of total power and device power, and a table of usage per month.