<p>Chaotic behaviors are widely present in numerous human-made and natural systems. The chaotic critical point is the key threshold parameter, at which these systems transition from ordered to disordered states. Such shifts often lead to catastrophic outcomes. It is imperative to design an effective method for predicting the chaotic critical point. However, most existing methods rely on prior knowledge of control parameters. In addition, they can only provide early warning signals but fail to quantitatively characterize the specific location where chaos occurs. To address these challenges, we propose a deep learning framework with the temporal attention mechanism to learning the dynamics rules underlying the time series of state variable for two tasks: 1) The classification task– whether the chaos will occur or not, 2) The prediction task– where the chaotic critical point occurs. We train the model on the classic chaotic system, the logistic map, and test its generalization on other chaotic systems. The experiment results show that the proposed method classifies the time series with the accuracy 98%, exhibits relative prediction inaccuracy less than 0.5% from ground truth, outperforms both the traditional and the deep learning methods. Moreover, our method is robustness with different noises, data lengths and variation rates, etc..</p>