Comparative modeling of long term CO 2 emissions in Indonesia’s power sector towards net zero

Abstract Indonesia’s power sector is a major contributor to CO 2 emissions, requiring precise long-term forecasting models to guide policy decisions. This study evaluates the effectiveness of Multiple Linear Regression (MLR), Polynomial Regression (PR), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) in predicting CO 2 emissions from 2014 to 2050. Key independent variables include GDP from the electricity and gas sector, total population, power plant capacity, electricity consumption per capita, and the share of coal in electricity generation. The findings reveal that LSTM projects the highest emissions at 1,069.88 MtCO 2 by 2050, followed by PR at 944.88 MtCO 2 , while GRU and MLR provide more conservative estimates at 697.47 MtCO 2 and 590.24 MtCO 2 , respectively. Model evaluation using R 2 , RMSE, MAE, MSE, and MAPE indicates that PR is the most accurate, with an R 2 of 0.984 and the lowest error values. In contrast, LSTM exhibits the highest errors, likely due to overfitting or limited training data. These results suggest that while regression models provide more stable projections, deep learning models capture complex patterns but may require larger datasets for better performance. Future research should explore hybrid modeling approaches that integrate statistical and deep learning techniques while incorporating additional factors such as carbon pricing policies and renewable energy expansion to improve forecasting accuracy and support Indonesia’s transition to a low-carbon energy future.