DATA ANALYSIS AND PREDICTION OF THE USE OF ELECTRIC ENERGIES FOR BUILDINGS USING ARTIFICIAL NEURAL NETWORK AND SUPPORT VECTOR MACHINE METHODS

• Main Author: TAUFIK HIDAYAT (NIM : 13311031) - IBNU WIDYATMOKO (NIM : 13311057), MUHAMMAD
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/23347
• Institution: Institut Teknologi Bandung

The issues of electric energy in Indonesia become one of the things that has not yet been solved until today and if it is left unchanged, it can lead to electric energy supply crisis. One of the ways to handle this crisis is by analyzing and predicting the use of electric energy. In this research, we will make a monitoring system of <br /> <br /> electric energy which is use in TP Rachmat Institut Teknologi Bandung building. In this system, we used control architecture and remote monitoring (Supervisory <br /> <br /> Control and Data Acquisition) so the system can be observed in anywhere and also be kept safely in cloud server. The data analysis is used to predict the meter <br /> <br /> of electric energy and to identify the element of the largest use based on a component’s custom at certain times. The method used to predict the use of energy is Artificial Neural Network (ANN) method, Backpropagation, and Support Vector Machine (SVM). These methods can make a predictor of the electric energy use by a customized training process. Input for both methods is a day and month code in 2014 with a target which is in the form of class that is suitable with the standard of Energy Consumption Intensity for the education buildings. Subsequently, the data in 2015 is used as a test data from created methods. The predictor design is made for two needs, which are daily data prediction and half-daily data prediction. ANN is designed by using backpropagation's algorithm with a hidden layer containing two neurons. The optimization result in trial and error way gives an activation function for hidden layer in the form of tansig, <br /> <br /> output layer in the form of purelin, and a training function with traingdy. A multiclass SVM is made by using algorithm of Error Correcting Output Code (ECOC). The optimization result shows an output of karnel polynomial that is used in every design. The value of Mean Square Error (MSE) becomes an error reference for testing result with the test data. The mistaken outcomes of ANN prediction in daily data on February, March, and April in sequence are 0.3636, 0.6818, and 0.5833. For half-daily data, the mistaken results of ANN are 0.9090, 0.6818, and <br /> <br /> 0.1667. Then, for the incorrect result of SVM on February, March, and April in sequentially are 0.1818, 0.7727, and 0.5833. For half-daily data, the mistaken results of SVM are 0.0565, 0.06818, and 0.4167. The deviation of MSE value <br /> <br /> between ANN method and SVM method is no more than 0.25, which means that the mistaken prediction does not reach one class so that both method are not much different.