DEVELOPMENT OF DIGITAL MODEL FOR ELECTRICITY CONSUMPTION PREDICTION IN SMART MICROGRID SYSTEMS USING DEEP LEARNING METHODS

• Main Author: Qanita, Adzra
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75391
• Institution: Institut Teknologi Bandung

The use of smart microgrids (MG) still faces several challenges, such as performance degradation caused by unstable environmental conditions or errors in system operation. The application of Micro Grid Digital Twin (MGDT) in smart MGs system makes it possible to analyze and optimize MG performance by utilizing data generated from MGDT on a digital twin (DT) platform. In the context of smart MGs, the ability of digital models to predict MG performance, especially in modeling electrical loads, is crucial. The pattern of electricity consumption in smart MG tends to vary, which makes it difficult for the digital model to reflect the actual conditions of the existing electrical load. Accurate predictions can assist in identifying electricity demand patterns, consumption trends, and significant load changes so that resource management at MG is more efficient. In this research, a digital model was developed to predict electricity consumption in smart MG systems at the Energy Management Laboratory, Engineering Physics, Bandung Institute of Technology. The digital model development pipeline uses the Cross Industry Standard for Data Mining (CRISP-DM) framework that engages historical data and deep learning. The modeling is done using several training data scenarios that consider the pattern of electricity consumption. The evaluation results show that modeling with a 12-month energy consumption pattern has the best performance. The deep learning used is Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM). The evaluation results for each parameter show that the RNN model has better accuracy than the DNN and LSTM models. The MAE, RMSE, and MAPE values generated by the RNN model when tested with test data are respectively 46.37 W; 68.33 W; and 5.85%. The results show that the developed model has better performance than the existing DNN model with a performance increase of 16.07% on MAPE.