Adaptive deep meta-learning ensembles for robust daily carbon emission forecasting in major continent economies

Carbon dioxide (CO₂) emissions are becoming an increasing concern for major countries worldwide, especially in communities across various scales. This situation emphasizes the critical importance of accurate carbon emission forecasting, particularly in effectively developing and adjusting short-term carbon reduction strategies. Nevertheless, daily-level forecasting is specifically challenging in a big data context due to carbon emission data's complicated, non-stationary, and non-linear character. To tackle this issue, an adaptive meta-learning framework for Deep ensemble architectures (AMLDE) model is developed in this research to forecast the CO₂ emission. Moreover, a robust optimization method called the ADL-TLBO (Adaptive teaching-learning-based optimization with differential evolution) algorithm is developed to extract the main parameters of the AMLDE model. To increase the forecasting accuracy of the AMLDE method, a hybrid decomposition method based on the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method, multi-scale complexity fusion (MCF), and variational mode decomposition (VMD) is developed. The proposed framework is applied to forecast CO₂ in three countries (China, the United States of America (USA), and India). The statistical results show that the AMLDE can predict CO₂ with high accuracy (correlation coefficient ( R ) up to 0.988) and low extreme error (RMSE = 0.168 to 0.382, and MaxAE = 0.836 to 2.516). Finally, based on risk analysis, the proposed model can achieve the "Very Low" risk to forecast CO₂ in all countries. Therefore, it can be used as a valuable forecasting tool to better manage the challenges of CO₂ emissions in countries.