•  
  •  
 

Corresponding Author ORCID Identifer

Keat Soon Tioh https://orcid.org/0009-0001-2514-2035

Section

Regular Issue

Abstract

This study evaluates the forecasting performance of three models of Autoregressive Integrated Moving Average (ARIMA), Prophet, and Bi-directional Long Short-Term Memory (BiLSTM) in predicting CO₂ emissions from the electricity grid in the United States. Using a real dataset from the California Independent System Operator (CAISO) that records CO₂ emissions at five-minute intervals between 2018 and 2020, the data underwent preprocessing steps including outlier removal, missing value imputation, and feature selection before model training. ARIMA parameters were determined using statistical tests such as ACF, Partial Autocorrelation Function (PACF), and Augmented Dickey Fuller (ADF), while Prophet addressed seasonality and calendar effects. BiLSTM, a deep learning model, was applied to capture complex temporal dependencies and non-linear relationships in the dataset. The models were evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics. The results show that BiLSTM achieved the lowest error values (MAE: 266.726; RMSE: 429.228), outperforming ARIMA (MAE: 1728.797; RMSE: 2203.464) and Prophet (MAE: 831.346; RMSE: 1102.884). These findings indicate that BiLSTM is more effective in modeling and forecasting highly variable and long-term CO₂ emission data. The study provides valuable insights for researchers and energy sector practitioners by demonstrating that machine learning-based forecasting can support carbon reduction initiatives and improve decision-making for electricity grid management. Moreover, the inclusion of a near-real-time alert system and dashboard design enhances the study’s practical relevance. Despite being limited to U.S. grid data, this research underscores the potential of BiLSTM models in achieving environmentally sustainable energy operations through accurate and timely CO₂ emission forecasting.

Download

Share

COinS