The Models

There are multiple models available to be created using ProLoaF.

Recurrent Neural Networks (RNN)

The default recommended model is an long-short-term-memory (LSTM). This model also allows the use of other recurrent model like gated-recurrent-units (GRU). The models are dynamically imported in the constructor so any model that implements torch.nn.Module and follows the signature of LSTM and GRU can be used as core model.

Transformer Models

A Transformer model is available for forecasting. This model could not show superior results when compared to RNN models while using a higher amount of resources and training time. Since theses test were somewhat limited the model could be reevaluated for a specific use-case, specifically when complex temporal dependencies are suspected and the size of the training dataset is sufficently large.

AutoEncoder Hybrid Model

This is an experimental model not recommended for practial use. This model is based on the RNN model, but combines the forecast with a decoder for recustruction of the input data during training, as would be the case in an autoencoder.