DE-Optimized Hybrid ARIMA-LSTM for Long Term Electricity Load Forecasting

Accurate long-term electricity load forecasting was essential for efficient energy planning and infrastructure development. This study addressed forecasting challenges in rapidly growing regions, such as East Java, where electricity demand was influenced by both linear and non-linear patterns. Conventional forecasting models, such as the Autoregressive Integrated Moving Average (ARIMA) effectively captured linear trends but failed to model non-linear dynamics, whereas networks with Long Short-Term Memory (LSTM) excelled with non-linear data but were often less effective when used alone. This research developed and evaluated an ARIMA-LSTM hybrid model optimized with the Differential Evolution (DE) algorithm to forecast electricity load until 2026. The model was trained and validated using historical daily load data from 2021 to 2023 from PT PLN UP2B East Java. This hybrid methodology first used ARIMA to model the linear components of the time series. The resulting residual errors, which contained non-linear patterns, were then modeled using an LSTM network. The DE algorithm was used to automatically optimize hyperparameters for both the ARIMA (p, d, q) and LSTM (units, learning rate, drop out etc.) components. The suggested hybrid model's performance was contrasted with that of the independent LSTM and ARIMA models. The results showed that the DE-optimized hybrid model achieved higher accuracy, yielding a Mean Absolute Percentage Error (MAPE) of 3.97 %, which was significantly better than the ARIMA model (12.39 % MAPE) and the LSTM model (4.50 % MAPE) on the validation set. According to these results, the suggested hybrid model was a dependable and extremely accurate instrument for predicting long-term loads, offering a solid basis for strategic energy planning.