Abstract: The traditional cascade prediction models do not consider the dynamics features in the process of information diffusion and rely on artificial marking features heavily, which have the problems of poor generalization and low prediction accuracy. This paper proposes a cascade prediction model (information cascade with dynamic graphs representation and self-attention, DySatCas) that combines dynamic graph representation and self-attention mechanism. The model adopts an end-to-end approach, avoiding the difficult problem of cascade graph representation caused by artificial labeling features, capturing the dynamic evolution process of cascade graphs through sub-graph sampling. And it introduces a self-attention mechanism to better integrate in the dynamic structure changes and temporal characteristics of the information cascade graph learned in the observation window, which can allocate weight values to the network reasonably, reduce the loss of information, and improve the prediction performance. Experimental results show that DySatCas has significantly improved prediction accuracy compared with the existing baseline prediction model.