Abstract: We propose a Physics-Enhanced Graph Attention Network (PE-GAT) to mitigate the prediction instability and physical inconsistencies prevalent in power grids with high renewable penetration.