Causal discovery in groups of time series is crucial for understanding complex systems, such as brain networks and climate dynamics. Existing methods often oversimplify group structures or neglect key intra- and inter-group temporal interactions, which limits the accuracy and depth of causal insights. To overcome these limitations, we propose CausalKGR, a novel framework that learns expressive latent representations specifically designed for Granger causal discovery across multiple time-series groups. The core idea of CausalKGR is to distil the temporal knowledge of each group into a high-dimensional, compact representation that serves as a condition for knowledge fusion and subsequent causal inference. We introduce a novel knowledge-conditional attention mechanism to selectively enhance the expression of key causal features within and across these learned group representations. The framework employs a two-stage self-supervised learning strategy: the first stage consolidates intra-group dynamics and inter-group interactions to generate robust group representations, and the second stage performs knowledge-conditional Granger causal discovery. Extensive experiments on synthetic and real-world datasets demonstrate that CausalKGR consistently outperforms existing methods, validating its robustness and effectiveness in complex group-level system analysis.