Far-side helioseismology is the branch of astrophysics dedicated to inferring the activity on the far hemisphere of the Sun using the wave-field from the near-side surface. Recently, the neural network FarNet, with a U-net architecture widely used for semantic segmentation, has been proven to improve the activity predictions of phase-sensitive holography, the standard method for applying far-side helioseismology. This was achieved using as inputs temporal sequences of nearside phase-shift maps. The network returned far-side activity probability maps for the central date of each input. We have developed FarNet-II, a new network that expands the capacities of FarNet by adding Convolutional LSTM modules and attention mechanisms to the model. This new tool uses the same phase-shift maps as inputs, but returns one activity probability map for each image date on the input sequence. We found that the prediction capabilities are greatly improved with respect to FarNet. Also, the outputs of this network keep better temporal coherence among them than those obtained with FarNet. Improved predictions from FarNet-II can contribute to a great number of applications on space weather, such as spectral irradiance and solar wind forecasting.