Research On Prediction Of Carbon Dioxide Emissions Based On Space-time Characteristics

Global warming caused by excessive carbon dioxide emissions will inevitably damage ecosystems,biodiversity and human economic activities.The deteriorating human living environment makes countries realize that it is urgent to take measures to reduce carbon emissions as soon as possible.Energy conservation and emission reduction has become an important task for all countries in the world.To reduce and control carbon dioxide emissions,it is very important to accurately predict carbon emissions.Accurate prediction of future carbon dioxide emissions will help to formulate more effective environmental policies,which may include energy policies,emission limitation policies,carbon trading policies,etc.The prediction model of carbon dioxide emissions can be roughly divided into traditional time series model,regression model and machine learning AI model.The traditional model is affected by the historical model and structural stability in practical application,and is not robust enough to predict nonlinear carbon dioxide emission data.Although machine learning model can deal with nonlinear data,carbon emission prediction is a complex problem,and a single model is difficult to apply to changing scenarios.At the same time,most prediction models regard carbon dioxide emission units as homogeneous individuals and independent of each other,but ignore the spatial connection and correlation between each unit.In recent years,the spatial effects of carbon dioxide emissions have attracted the attention and attention of researchers,and the spatial factors have gradually been included in the analysis of carbon dioxide emission impact factors.However,no research has applied the impact of spatial effects to the prediction model of carbon dioxide emissions.In this paper,GCN and GAT graph neural networks are used to learn the spatial characteristics of carbon emissions between regions.Based on the LSTM model,two spatiotemporal prediction models,LSTM-GCN and LSTM-GAT,are constructed.The similarity between the two models is that the spatial characteristics of carbon dioxide emissions are learned by using the graph neural network,and then the data with spatial characteristics are brought into the LSTM model to learn the characteristic information of time series.According to the research,similar carbon emission patterns and trends may still exist between non-adjacent regions.The spatial effect of carbon emissions is not only related to the geographical location of regions,but also can guide the prediction of carbon dioxide emissions among regions with similar carbon emission patterns.Therefore,this paper proposes a spatiotemporal fusion matrix that combines the spatial weight matrix and the carbon emission similarity matrix.It is introduced into the spatiotemporal prediction model as the adjacency matrix.This article reviews and summarizes the existing carbon emission prediction models,spatial effects,and spatiotemporal prediction related theoretical achievements,and selects carbon dioxide emission data from a total of 11 countries collected on the Carbon Monitor website from January 1,2020 to December 31,2022 for comparative experimental research,obtaining meaningful conclusions.Firstly,for the spatiotemporal prediction models LSTM-GCN and LSTM-GAT,both outperform other basic models and LSTM models in various evaluation indicators of prediction results.LSTM-GCN has improved LSTM by9.15%,4.13%,and 3.33% in MAPE,MAE,and RMSE,respectively.LSTM-GAT improved LSTM by 11.05%,5.83%,and 6.18% on MAPE,MAE,and RMSE,respectively.On the basis of LSTM-GAT,Pearson correlation coefficient,grey correlation degree,and dynamic time warping algorithm are used to further analyze the similarity of carbon emission time series in different regions.Three different spatiotemporal fusion matrices are introduced into the LSTM-GAT model to construct three improved models: LSTM-GAT-T1,LSTM-GAT-T2,and LSTM-GAT-T3.The research results indicate that the LSTM-GAT-T1 model achieved the best prediction results on three indicators: MAPE,MAE,and RMSE.At the same time,in order to study the robustness,stability,and long-term prediction ability of the model,multi-step prediction experiments were conducted in this paper.Finally,LSTM-GAT-T1 outperformed other basic models and spatiotemporal prediction models in all prediction ranges,indicating that the LSTM-GAT-T1 model can achieve the best prediction results in short-term and long-term carbon emissions prediction.Accurate carbon dioxide emission prediction provides basis and support for environmental pollution control,energy conservation and emission reduction plans,and plays a pivotal role in mitigating global warming.From the perspective of management,building a scientific carbon dioxide emission prediction model is of great significance for the relevant departments of governments to formulate management measures to improve carbon emissions in the medium and long term.From the study of the spatial impact of carbon dioxide emissions,we also get the following enlightenment: only by strengthening cooperation between countries can we achieve "carbon neutrality" at an early date.Global joint emission reduction requires the joint efforts of governments,enterprises and the public to achieve the goal of controlling carbon dioxide emissions and mitigate the threat of climate change to human society and the natural environment.