TLDR: This research paper presents an advanced ensemble model for forecasting CO2 emissions in Taiwan over the next decade (2023-2032). By comparing 21 univariate and multivariate time series models, the study identified Feedforward Neural Network (FFNN), Support Vector Regressor (SVR), and Random Forest Regressor (RFR) as top performers. These were then combined using a custom stacked generalization technique with Linear Regression, resulting in a highly accurate model (SMAPE of 1.407). The forecast projects a slight decline in Taiwan’s CO2 emissions by 2032, suggesting the effectiveness of current government policies.
A recent study delves into the critical issue of carbon dioxide (CO2) emissions in Taiwan, a region grappling with significant air pollution due to its high population density and heavy reliance on fossil fuels. The research, titled “Decade-long Emission Forecasting with an Ensemble Model in Taiwan,” provides a comprehensive analysis of various time series models to predict future emission trends, offering valuable insights for policymakers.
Taiwan’s energy sector, heavily dependent on fossil fuels like coal, oil, and natural gas, accounts for over 70 percent of its total annual CO2 emissions. In 2021, the island emitted over 257 million metric tons of CO2, a figure that has steadily increased over the past decade. This rise in emissions poses severe health risks, contributing to respiratory diseases, and leads to environmental degradation, affecting water and soil quality, and reducing biodiversity.
In response, the Taiwanese government has initiated several policies to curb emissions. These include updating the Nationally Determined Contribution (NDC) under the Paris Agreement in 2021, aiming for a 50 percent reduction by 2030 compared to 2005 levels. The Renewable Energy Development Act, amended in 2019, promotes solar and wind energy, and Taiwan has officially committed to achieving net-zero emissions by 2050. These efforts have already shown a slower rate of increase in total CO2 emissions.
The study compared 21 commonly used time series models, encompassing both univariate (single variable) and multivariate (multiple variables) approaches. The dataset, acquired from Our World in Data, included annual CO2 emissions per capita, total gas consumption, total coal consumption, and total oil consumption from 1965 to 2022. After rigorous data preprocessing, including ensuring stationarity and checking for multicollinearity, the data was split into training (1967-2012) and testing (2013-2022) sets.
The research found that multivariate models, which consider the interplay between CO2 emissions and factors like gas, coal, and oil consumption, significantly outperformed univariate models. Among these, Feedforward Neural Network (FFNN), Support Vector Regressor (SVR), and Random Forest Regressor (RFR) emerged as the top performers. To further enhance prediction accuracy and robustness, a novel ensemble model was developed. This model integrated the predictions of FFNN, SVR, and RFR using Linear Regression as a meta-model, strategically assigning optimal weights to each base model.
This custom stacked generalization ensemble technique proved superior, achieving an impressive Symmetric Mean Absolute Percent Error (SMAPE) of 1.407, with no signs of overfitting. The ensemble model demonstrated a remarkable ability to predict even slight fluctuations in emission trends, making it highly reliable for long-term forecasting.
Using this optimized ensemble model, the study provides a decade-long emission projection for Taiwan, spanning from 2023 to 2032. The forecasts indicate a slight decline in CO2 emissions towards the end of this period. This projection offers promising evidence that the government’s current policies and measures are beginning to yield positive results in reducing CO2 emissions over time.
Also Read:
- Advancing Time Series Forecasting with Residual-Stacked Gaussian Linear Models
- Greener Clouds: Federated Learning Predicts Data Center Energy Use While Protecting Privacy
This research underscores the power of machine learning in forecasting complex environmental trends and offers a data-driven tool for policymakers to make more informed decisions in their ongoing efforts to combat climate change. For more detailed information, you can access the full research paper here.
