As climate change continues to detrimentally affect human lives, accurately projecting carbon dioxide (CO2) emissions, one of the largest contributors to climate change, is becoming increasingly critical. However, forecasting CO2 emissions in Taiwan has become challenging due to its rapid development. This paper presents a detailed study of state-of-the-art univariate and multivariate time series models and then proposes a novel hybrid ensemble model for accurate CO2 forecasting in Taiwan. We evaluated 10 univariate models—ARIMA, Holt-Winters, SARIMA, Auto ARIMA, Simple Exponential Smoothing, Holt's Linear Trend Model, Theta, FFT, and Naive Drift—and 11 multivariate time series models—DFM, LSTM, FFNN, VAR, BVAR, VECM, Ridge Regression, Elastic Net Regression, SVR, RFR, and Decision Tree. Our custom dataset, spanning from 1965 to 2022, includes annual data on CO2 emissions as well as gas, coal, and oil consumption. Using standard evaluation metrics, we identified the three top-performing models: Feedforward Neural Network (FFNN), Support Vector Regressor (SVR), and Random Forest Regressor (RFR). We then utilized stacked generalization to combine their predictions with a meta-model. This proposed hybrid ensemble model achieved a MAPE score of 1.398%, demonstrating superior and more robust performance compared to previously proposed models. After extensive optimizations, the model was employed to forecast CO2 emissions in Taiwan for the next 10 years. This study provides a novel hybrid ensemble model and a robust framework for forecasting CO2 emissions, assisting policymakers and industry leaders in making informed decisions to reduce CO2 emissions.