Figure 10.

The PreD-Net architecture. The PreD-Net used for the prediction of precursors, is designed to operate with and without lagged variables. The central branch acts as a convolutional autoencoder on the feature of the problem, generating a latent space of dimension $n_{l}x16$, where $n_l$ is the number of lagged variables. The GRU layers act on the latent space reconstructing the temporal dependencies between the $n_l$ lagged considered steps. At the same time, a dilated convolution extracts the relations between every single feature with its lagged versions. The three outputs are concatenated and fed to a dense sequence of layers to ensemble the extracted information. A softmax output returns the probability of each sample belonging to class 0 (background seismicity) or class 1 (precursors).