Abstract

Background: With the rapid economic growth and the accelerated process of industrialization, the production activities of enterprises within parks have significantly increased, leading to a continuous rise in carbon emissions. Under the context of the "dual carbon" goals, studying the prediction of carbon emissions in typical parks holds significant practical importance. It is not only a key measure to address climate change but also an important pathway to achieve sustainable development.

Objective: In order to predict the carbon emissions of the typical park more accurately, we propose a carbon emissions prediction model HA-SCINet.

Methods: The model uses a recursive downsampling-convolution-interaction architecture. In each layer, the long-term dependence in time series data is extracted by HyperAttention. Then through the L-layer SCI-Block of the binary tree structure, the down-sampling interactive learning extracts both short-term and long-term dependencies. These extracted features are merged and reorganized, and added to the original time series to generate a new sequence with enhanced predictability. Finally, employ a fully connected network for forecasting the enhanced sequence. The carbon emission data of the typical park serve as input, leading to higher accuracy prediction results through the Stacked K-layer stacked HA-SCINet.

Results: The mean square error (MSE) and mean absolute error (MAE) of HA-SCINet prediction model are 0.0819 and 0.204 respectively, outperforming the mainstream Dlinear and Nlinear models.

Conclusion: The experimental results show that the devised model outperforms in predicting carbon emissions, and is better suited for forecasting carbon emissions within the context of the typical park.

Keywords: Carbon emission prediction, HA-SCINet, SCInet, mean square error (MSE), mean absolute error (MAE), Stochastic Impacts by Regression Population, Affluence, and Technology (STIRPAT)..