Application of Time Series Regression, Double Seasonal ARIMA, and Long Short-Term Memory for Short-Term Electricity Load Forecasting

Electricity load must be accurately estimated since electricity is non-storable. If electricity is generated more than customer’s demand, it will be wasted, and the power of generator should be lowered. Meanwhile, if electricity is generated less than customer’s demand, it may cause power outage and must undergo backup-plant operation. This study compares three forecasting methods such as time series regression (TSR), double seasonal autoregressive integrated moving average (DSARIMA), and long short-term memory (LSTM) to predict electricity load. These three methods can be applied to forecast electricity load which has double seasonal pattern. Each method has an ability to capture the data pattern. The observation data is half-hourly recorded electricity load of East Java Province in Mega Watt (MW) units from January 1^st2020 to January 31^st, 2023. This study obtains the out-sample symmetric mean absolute percentage error (sMAPE) of those methods as many as 0.6436%, 0.5504%, and 0.9713%, respectively. From these findings, DSARIMA ([2, 10, 11, 12, 15, 16, 17, 18, 19, 20, 21, 22, 23], 1, [1, 2, 3, 7, 8, 30, 34, 35, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48]) (0, 1, 1)⁴⁸(0, 1, 1)³³⁶is apparent to be the best forecasting model for its lowest out-sample sMAPE value. This model gained a white noise residual, but it does not have a normal distribution due to outliers.