How to add a model

Model Structure

All models are stored in the workspace models folder as .py files. Each model defines two functions: one for training and one for forecasting. Cross-validation, evaluation, and plotting are handled by the core engine, so model files only need to implement model-specific logic.

Training Function

The training function trains the model using the training set predictors and target values. It returns a single object, model, which contains the trained model.

Forecast Function

The forecast function uses the trained model along with the training and testing predictors. It returns two outputs: forecasts for the training and testing sets.

How to Add a Model

To add a new model, follow these steps:

1. Create a new model file in your workspace

Create a new .py file under your workspace models folder. The file name must start with a model ID prefix, for example m19_my_model.py. The functions inside must match the file stem:

• train_model_m19_my_model(...)
• produce_forecast_m19_my_model(...)

You can start from the template file models/mXX_template.py in your workspace.

If the model uses randomness, use a standard seed-like hyperparameter key such as seed, random_seed, or random_state. PyNNLF can override these keys from the central reproducibility.seed value in specs/pynnlf_config.yaml so experiments remain comparable.

2. Add hyperparameters for the new model

Add a new top-level key in example_project/models/hyperparameters.yaml. The key must match the full model file stem, for example m19_my_model.

m19_my_model:
  hp1:
    bias: 0.0
  hp2:
    bias: 5.0

3. Select the model in your experiment spec

Set the model in example_project/specs/experiment.yaml:

model: m19
hyperparameter: hp1

The model will be discovered automatically based on the m19_*.py file in your workspace models folder. No dispatch code changes are required.