PREDICTION MODELING OF ELECTRICITY CONSUMPTION IN HOUSEHOLD CUSTOMER SEGMENTS USING MACHINE LEARNING : CASE STUDY PT PLN (PERSERO)

• Main Author: Faeda Insani, Akhmad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86673
• Institution: Institut Teknologi Bandung

Indonesia's rapid economic growth has driven a significant increase in electricity consumption across various sectors, including households, businesses, and industries. The continuously growing demand for electricity poses challenges for the energy sector to maintain a reliable and efficient supply, particularly in meeting the needs of households as one of the largest consumer segments. This study aims to develop a model for predicting electricity consumption in the household segment using the Long Short-Term Memory (LSTM) algorithm, which is known for its superior performance in analyzing time series data, especially with a multivariate approach. Considering the complexity of factors influencing electricity consumption, this research focuses on identifying the most relevant sliding window configurations and variable combinations to improve prediction accuracy. This study employs six primary independent variables: electricity consumption, customer growth, additional connected capacity, inflation rate, Gross Domestic Product (GDP) of household consumption, and average temperature. Historical data spanning from 2004 to 2023 was collected from various official sources, including PT PLN (Persero), the Central Statistics Agency (BPS), Bank Indonesia (BI), and the Climate Change Knowledge Portal. The analysis begins with an exploration of correlations between variables to determine their relevance to household electricity consumption. The data is then processed to train the LSTM model to predict electricity consumption based on different sliding window scenarios: 1 month, 3 months, 6 months, and 12 months. The findings reveal that the variable combinations selected based on initial correlation analysis significantly influence the prediction model's accuracy. Among the four sliding window scenarios tested, the 12-month window size scenario yielded the best results. The variable combination used in this scenario includes customer growth, inflation rate, GDP of household consumption, and household electricity consumption. With this configuration, the LSTM model achieved a Mean Absolute Percentage Error (MAPE) of 2.44%, demonstrating a very high level of prediction accuracy. However, the model tends to be less effective in capturing monthly fluctuation patterns. This study also provides critical insights into the impact of window size on the model's ability to identify seasonal patterns or long-term trends in electricity consumption data. The contribution of this research lies not only in developing an accurate prediction model but also in its implications for future energy consumption planning and management. The proposed model can be used as a tool to support strategic decision-making by both the government and energy providers to ensure an adequate electricity supply for the household segment. Furthermore, these findings open opportunities for similar research development in other tariff segments, such as business, industry, social, and public sectors. Thus, this study not only enriches academic literature but also delivers practical benefits in supporting the transition toward a more sustainable energy system in Indonesia.