Annual CO2 Emissions Forecasting in Vietnam Using a PCA-Stacked Bidirectional LSTM Model

Vietnam’s rapid economic growth has brought substantial improvements in living standards, industrial capacity, and infrastructure. However, it has also resulted in significant environmental stress, primarily driven by increased energy consumption for production and consumption activities. This study proposes a principal component analysis (PCA)–stacked bidirectional long short-term memory (PCA-stacked BiLSTM) framework, evaluated using an expanding-window forecasting strategy. Drawing on a complex adaptive systems (CAS) perspective, PCA is used to transform highly interdependent macroeconomic predictors into latent components, thereby improving model stability. The expanding-window approach ensures that forecasts are generated sequentially using only historical information, thereby providing a realistic assessment of out-of-sample performance. With annual data spanning 1990–2023 for Vietnam, the proposed model is compared with benchmark methods, including ARIMA, support vector regression, KNN, and unidirectional LSTM models. The results indicate that while classical time-series models remain highly competitive in small-sample settings, the stacked BiLSTM improves forecasting robustness relative to standard LSTM architectures and machine learning baselines. This study contributes a methodologically grounded framework that integrates complexity theory, multicollinearity mitigation, and realistic time-series validation, offering more reliable long-horizon CO2 emissions forecasts in data-constrained environments.