Prediction of Electrical Load at PT PLN (Persero) Jayapura Using Recurrent Neural Network and Long Short-Term Memory Models

Abstract

 The primary challenge in Power System Operation Planning is the uncertainty in predicting electricity load. Inaccurate electricity demand forecasts can lead to issues such as resource wastage, increased operational costs, and supply failure risks. Traditionally, operational planning has relied on estimating load history using Microsoft Excel worksheets, with calculations based on load
growth (%) from previous periods. This research aims to improve the accuracy of electricity load prediction for system operation planning at PT. PLN (Persero) Jayapura by utilizing deep learning models, specifically Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM). For model optimization, a grid search method was employed for hyperparameter tuning, ensuring the best performance in load forecasting. The study was conducted at PT. PLN (Persero) Jayapura, using daily electricity load data from January 2020 to August 2024, sourced from the SCADA (Supervisory Control and Data Acquisition) histori server. The results showed that the LSTM model outperformed the traditional RNN. While the RNN model achieved a Mean Absolute Error (MAE) of 1.106, a
Root Mean Squared Error (RMSE) of 1.7650, and a Mean Absolute Percentage Error (MAPE) of 0.0142, the LSTM model demonstrated more accurate predictions with a MAE of 1.0047, RMSE of 1.6186, and MAPE of 0.0129. These findings demonstrate the potential of LSTM, enhanced by grid search optimization, for improving load forecasting accuracy and contributing to more reliable power system operation planning.