*2.2. Connectivity Models*

The ANNs evaluated in this study were selected based on their suitability for time series analysis (Table 1). While TCDF outputs attention scores, for which higher scores are used to represent stronger connectivities, LSTM-NUE makes use of Granger Causality scores equaling NNGC = errreduced−errfull, which are then binarized [18]. Conv2D uses the R2-score between the real and predicted values of the current target (i.e., the time steps to be predicted). While using TRGC as implemented by [1], only binary GC scores are outputted. The configuration (i.e., the used parameters) of each ANN was determined using a data-driven approach, such that for each ANN, the parameters returning the best results were chosen. This parameters pre-testing was performed with different simulated data sets (i.e., differing from the data sets that were used to report the final results).

**Table 1.** Set-up of the compared ANNs and TRGC.

