ANALYSYS OF LONG SHORT-TERM MEMORY (LSTM) AND WAVELET DECOMPOSITION APPROACH FOR DEMAND FORECASTING OF DISTRIBUTION TRANSFORMER: STUDY CASE PLN UP3 BALIKPAPAN

• Main Author: Bagaswara, Tito
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86906
• Institution: Institut Teknologi Bandung

In the electrical system, distribution transformers play a critical role in transmitting electrical energy from high voltage to low voltage according to customer needs. Accurate planning of distribution transformer requirements is crucial, considering the dynamic and fluctuating energy consumption patterns. This study aims to develop a predictive model for distribution transformer requirements based on the Long Short-Term Memory (LSTM) algorithm, both with and without wavelet decomposition integration, to improve prediction accuracy. Historical data on distribution transformer usage from 2013–2023, obtained from the SAP ERP application, was analyzed using a quantitative approach and stationarity tests. The LSTM model was employed to predict transformer needs across various capacities (50 kVA, 100 kVA, 160 kVA, and 250 kVA). Wavelet decomposition was implemented to separate the main trend components (lowfrequency) from fluctuations (high-frequency) to reduce noise in the data. The model evaluation was conducted using Mean Absolute Error (MAE), Relative Mean Absolute Error (RelMAE), and Symmetric Mean Absolute Percentage Error (SMAPE) metrics. The research results show that the integration of wavelet decomposition into LSTM significantly improves accuracy, with the best MAE value of 3,00 achieved for the 250 kVA transformer capacity. The Stacked LSTM model demonstrated superior performance for the 50 kVA (MAE 1,51) and 100 kVA (MAE 3,19) capacities, while the Simple LSTM model provided the best results for the 160 kVA capacity (MAE 2,14). Evaluation using RelMAE indicates that the predictive model achieves error levels lower than the data's standard deviation, with values of 0,61 (250 kVA), 0,63 (100 kVA), 0,69 (160 kVA), and 0,89 (50 kVA). This confirms that the model is capable of producing reliable predictions to address transformer shortages at PLN UP3 Balikpapan.