Interpretable short-term carbon dioxide emissions forecasting based on flexible two-stage decomposition and temporal fusion transformers

Abstract

Carbon emissions play a pivotal role in exacerbating the global warming crisis and driving climate change. Accurate and consistent projections of carbon emissions are of utmost importance for nations worldwide, as they shape emission reduction strategies and expedite the pursuit of carbon peaking and carbon neutrality goals. While previous studies have focused on hybrid methodologies for carbon emission forecasting, yielding commendable predictive performance, these approaches often overlook the significance of internal interpretability within the forecasting models. In light of this gap, this study introduces a groundbreaking and elucidating hybrid carbon emission forecasting model that amalgamates the two-stage layer decomposition method, adaptive differential evolution with optional external archive (JADE), and temporal fusion transformers (TFT). To begin with, a series of sub-sequences is derived by employing a flexible two-stage decomposition strategy, which leverages a linear-nonlinear decomposition criterion to thoroughly extract the fluctuating characteristics inherent in the carbon emission series. Subsequently, the JADE algorithm intelligently and efficiently optimizes the parameter combinations within the TFT model, ensuring both stability and reliability of the prediction framework. Empirical investigations conclusively demonstrate the remarkable applicability and efficacy of the proposed model in short-term carbon emission forecasting. By delving into the interpretability of the model's results, the study enhances the capacity of policymakers to devise well-informed strategies based on comprehensive insights gleaned from the forecasting process.