LOAD FORECASTING FOR DAILY LOAD OPERATIONAL PLAN USING LSTM (CASE STUDY: SOUTH SULAWESI SUB SYSTEM)

Abstract

The electrical load required in an electricity sub-system changes every day. Electric power operators must be able to generate and distribute electricity according to consumer needs. In the Sulawesi sub-system, the power plants used are still dominated by fossil fuel generators, so that in their operations, fuel requirements need to be given serious attention. Planning a good daily electricity consumption is needed so that the fuel cost becomes optimal. In the current condition, the load forecasting for the Daily Load Operation Plan (ROH) is still based on Expert Judgment, which is different for each forecaster. With a fairly large error tolerance limit of 4%. We need a load forecasting instrument capable of better error tolerance. Forecasting methods such as ARIMA, SARIMA and ARIMAX have been used for many years. In recent years, several artificial intelligence techniques such as Neural Network and machine learning have been developed for time series analysis. And recently, more accurate forecasting results are shown by Artificial Neural Network (ANN) and Recurrent Neural Network (RNN) compared to traditional forecasting methods. Long Short Term Memory (LSTM) is a model of RNN that uses past data (Long Term) to predict current data (Short Term). Electric load in Sulawesi subsystem used as data training after normalized using min-max normalization. The LSTM model is made with different data input. Forecasting performance of each model is then evaluated based on the RMSE and MAPE values. Of the several data input models, forecasting models with daily data input show better performance than other scenarios. The MAPE and RMSE values obtained were 2.384% and 33.95, respectively.