Carbon Emission Factor Multi-Time Scale Prediction with Adaptive Graph Convolution Strategy

The incremental attainment of the dual carbon objective has prompted a growing emphasis on precision carbon management as a significant area of concern. The carbon emission factor serves as a connection between electricity consumption and carbon emissions, facilitating analysis and management for grid operators. However, its complex characteristics make it challenging to predict, hindering grid managers in planning future electricity consumption. To address this issue, we introduce CEFPNet, a framework designed to obtain the prediction result of the carbon emission factor by extracting correlations between variables and temporal characteristics at different time scales. In CEFPNet, the multivariate time series (MTS) input first exposes its features through the embedding layer and then enters multiple G2CBlocks with residual connections to extract features and obtain prediction results. Within each G2CBlock, we apply the Fast Fourier Transform (FFT) to identify different time scales, segment the sequence into a three-dimensional tensor, and perform graph convolution to capture inter-variable relationships. Subsequently, multiple two-dimensional convolutions are applied to extract time series information, which is then reshaped back into a two-dimensional tensor to produce the final prediction. Results from three real datasets show that CEFPNet performs better than comparison methods in predicting the carbon emission factor. In addition, ablation experiments are delivered to prove the effectiveness of CEFPNet's substructures.