As the inertia is unevenly distributed in the system, the spatial and temporal dynamic phenomena for the frequency are obvious after a disturbance. After the disturbance, accurately obtaining the dynamic frequency nadir can quickly formulate the corresponding frequency stability control measures to prevent system frequency collapse. Here, a novel convolutional long-short term memory mechanism and attention network combined model (ATT-ConvLSTM) is used to predict the frequency nadir of each generator, which can extract the spatial-temporal correlations among the sampled data. Faced with noise interference, the attention mechanism can effectively identify input features from the sampled data to improve the online application performance. The simulation experiments are performed in an improved New England 10-machine 39-bus system and an actual complex system. Over existing models, the empirical simulation analysis shows the superiority and applicability of the combined model.