A GNN-based Day Ahead Carbon Intensity Forecasting Model for Cross-Border Power Grids

Carbon intensity forecasting of power grids is critical to the optimization of demand-side consumers. Recently, cross-border power grids have emerged, i.e., those allowing electricity to be transmitted across different national transmission systems. Cross-border power grids substantially increase the sharing of highly variable renewable energy sources (VRE), leading to greater economic benefits and increased reliability. In Europe, the total volume of cross-border electricity that is exchanged comprises 13% of the annual net electricity that is generated. Current studies on carbon intensity forecasting, however, apply to individual regional power grids. In cross-border grids, the carbon intensity of a regional grid depends not only on that of its own electricity but also on the carbon intensity from the electricity exchanged with cross-border grids. Thus, if the cross-border electricity exchange is not captured appropriately, significant forecasting errors can occur. In this paper, we formulate a new Carbon Intensity Forecasting for Cross-border Grids (CFCG) problem by proposing and integrating carbon flows generated by cross-border electricity exchanges. The challenge is to capture the complex spatial and temporal dependencies that are involved. We propose a CFCG model based on a Graph Neural Network (GNN) submodel to learn the spatial dependencies and a Long Short Term Memory (LSTM) submodel to learn the temporal dependencies. We evaluate the CFCG model using real-world data from the cross-border power grids in Europe involving 28 member countries. We compare five baseline models. Our results show that the CFCG model achieves an average improvement of 26.46% or 20.34% as compared to state-of-the-art forecasting models based on regional grids or one-hop neighbor grids, respectively.