median_forecaster#

Median regressor based forecasting models for energy forecasting.

Provides median regression models for multi-quantile energy forecasting.

Note that this is an autoregressive model, meaning that it uses the previous predictions to predict the next value.

This regressor is good for predicting two types of signals:

  • Signals with very slow dynamics compared to the sampling rate, possibly with a lot of noise.

  • Signals that switch between two or more states, which are random in nature or depend on unknown features, but tend to be stable in each state. An example of this may be waste heat delivered from an industrial process. Using a median over the last few timesteps adds some hysteresis to avoid triggering on noise.

Tips for using this regressor:

  • Set the lags to be evenly spaced and at a frequency matching the frequency of the input data. For example, if the input data is at 15 minute intervals, set the lags to be at 15 minute intervals as well.

  • Use a small training dataset, since there are no actual parameters to train.

  • Set the frequency of the input data index to avoid inferring it. Inference might be a problem if we get very small chunks of data in training or validation sets.

  • Use only one training horizon, since the regressor will use the same lags for all training horizons.

Classes#

MedianForecaster(**data)

Median forecaster using lag features for predictions.

MedianForecasterHyperParams(**data)

Hyperparameter configuration for median forecaster.