A Study on Provincial-Level Carbon Emission Forecasting Based on Multimodal Data and Spatio-Temporal Neural Networks

Abstract

Against the backdrop of global warming and the ‘dual carbon’ targets, accurate forecasting of provincial carbon emissions is of great significance for formulating differentiated emission reduction policies and optimising energy structures. This study addresses issues such as the inability of traditional statistical models to capture non-linear characteristics, the limited predictive accuracy of single machine learning models, and the failure of existing deep learning methods to account for the spatial spillover effects of carbon emissions. The study first constructs a provincial multi-modal dataset integrating features such as GDP, industrial output, energy prices and policy intensity, and establishes a spatial adjacency matrix based on geographical proximity; secondly, it proposes a hybrid GCRN-GCNLSTM architecture that accounts for the spatio-temporal dependencies of carbon emissions; finally, model interpretability is enhanced through feature importance analysis and residual diagnostics. Using six major Chinese provinces—Shandong, Guangdong, Jiangsu, Hebei, Zhejiang and Henan—as the empirical subjects, the study conducted a comparative evaluation of ARIMA, LSTM, GCN-LSTM and ensemble models. The results indicate that the LSTM model performed best on the test set, with a coefficient of determination (R²) of 0.8383, a mean absolute error (MAE) of 0.4191 million tonnes, and a root mean square error (RMSE) of 0.5102 million tonnes; the mean absolute percentage error (MAPE) of the ARIMA model was 22.01%, indicating a significant linear component in the carbon emissions time series; The GCN-LSTM model, which incorporates geospatial information, did not outperform the LSTM model. Analysis suggests this may be attributable to the time lag associated with spatial spillover effects, an inappropriate graph structure design, and spatial redundancy in the economic features. The contribution of this study lies in proposing a ‘multimodal features + time series modelling’ forecasting paradigm, systematically evaluating the applicability of various models in provincial carbon emissions forecasting, and providing data support and theoretical grounds for improvements in spatial modelling and the formulation of differentiated emission reduction policies.