Methods

All eight reference methods live under swissrivernetwork/benchmark/model.py and are wired up through swissrivernetwork/benchmark/training.py.

method	Temporal backbone	Graph backbone	Typical use
lstm	LSTM	—	Per-station baseline
graphlet	LSTM	LSTM on neighbor predictions	Two-step graph variant
lstm_embedding	LSTM	station embedding	Shared per-graph model
stgnn	LSTM	PyG GNN	Joint spatio-temporal
transformer	Transformer	—	Isolated Transformer
transformer_graphlet	Transformer	same as graphlet	Graph variant of Transformer
transformer_embedding	Transformer	station embedding	Shared Transformer
transformer_stgnn	Transformer	PyG GNN	Joint Transformer ST-GNN

Transformer positional encodings

Each Transformer variant can use one of:

• sinusoidal
• learnable
• rope

Set with -pe on the CLI or via config["positional_encoding"] inside the __main__ blocks of the training drivers.

Where hyperparameters live

• ray_tune.py — search space definitions. This is the source of truth for tuned runs.
• train_single_model.py / train_isolated_station.py — inline dicts in __main__ for manual / debug training.

See CLI reference for the mapping from command-line flags to config keys.