Precipitation nowcasting, a short-term weather forecasting technique for few hours, is vital for urban management and disaster prevention. While current approaches have advanced through data processing and deep learning architectures, yet their black-box nature limits understanding of in-ternal mechanisms and hinders targeted improvements. We propose Game GAN Nowcasting (G2N), an interpretable generative model that leverages Multi-order Interaction theory to decompose precipitation nowcasting into quantifiable feature interactions. G2N introduces a novel game-theoretic regularization term in the loss function that guides the model to learn physically meaningful feature representations. Our pixel-level interaction visualization method reveals the model’s decision-making process through intuitive heatmaps. Experiments on multiple benchmark datasets demonstrate that G2N achieves average improvements of 10.73%, 24.52%, and 5.80% in Critical Success Index (CSI), Heidke Skill Score (HSS), and Structural Similarity Index (SSIM), respectively, while providing clear insights into its prediction mechanisms. The enhanced interpretability enables better understanding of precipitation evolution patterns and facilitates model refinement for operational weather forecasting.