Forecaster

CINNForecaster

Conditional Invertible Neural Network (cINN) Forecaster.

This forecaster uses a cINN to forecast the time series. The cINN learns a bijective mapping between the time series and a normal distributed latent space. The latent space is then sampled and transformed back to the time series space. The cINN is conditioned on statistical and fourier term based features of the time series and the provided exogenous features. This forecaster was applied in the BigDEAL challenge by the KIT-IAI team and is described in [1]_.

Schnellstart

python

from sktime.forecasting.cinn import CINNForecaster

estimator = CINNForecaster(n_coupling_layers=10, hidden_dim_size=32, sample_dim=24, batch_size=64, encoded_cond_size=64, lr=0.0005, weight_decay=1e-05, sp_list=None, fourier_terms_list=None, window_size=720, num_epochs=50, verbose=False, f_statistic=None, init_param_f_statistic=None, deterministic=False, lag_feature='mean', patience=5, delta=0.0001, val_split=0.2)

Parameter(19)

n_coupling_layersint, optional (default=15): Number of coupling layers in the cINN.
hidden_dim_sizeint, optional (default=64): Number of hidden units in the subnet.
sample_dimint, optional (default=24): Dimension of the samples that the cINN is creating
batch_sizeint, optional (default=64): Batch size for the training.
encoded_cond_sizeint, optional (default=64): Dimension of the encoded condition.
lrfloat, optional (default=5e-4): Learning rate for the Adam optimizer.
weight_decayfloat, optional (default=1e-5): Weight decay for the Adam optimizer.
sp_listlist of int, optional (default=[24]): List of seasonal periods to use for the Fourier features.
fourier_terms_listlist of int, optional (default=[1, 1]): List of number of Fourier terms to use for the Fourier features.
window_sizeint, optional (default=24*30): Window size for calculating the rolling statistics using the WindowSummarizer.
lag_feature: str, optional (default=”mean”): The rolling statistic that the WindowSummarizer should calculate.
num_epochsint, optional (default=50): Number of epochs to train the cINN.
verbosebool, optional (default=False): Whether to print the training progress.
f_statisticfunction, optional (default=default_sine): Function to use for forecasting the rolling statistic.
init_param_f_statisticlist of float, optional (default=[1, 0, 0, 10, 1, 1]): Initial parameters for the f_statistic function.
deterministicbool, optional (default=False): Whether to use a deterministic or stochastic cINN. Note, deterministic should only used for testing.
patienceint, optional (default=5): Number of epochs to wait before stopping the training.
deltafloat, optional (default=0.0001): Minimum change in the validation loss to consider as an improvement.
val_splitfloat, optional (default=0.2): Fraction of the data to use for validation.

Beispiele

>>> from sktime.forecasting.cinn import CINNForecaster
>>> from sktime.datasets import load_airline
>>> y = load_airline ()
>>> model = CINNForecaster (window_size = 100)
>>> model. fit (y) CINNForecaster(
... )
>>> y_pred = model. predict (fh = [1, 2, 3 ])

Referenzen

..[1] Heidrich, B., Hertel, M., Neumann, O., Hagenmeyer, V., & Mikut, R.: (2023). Using conditional Invertible Neural Networks to Perform Mid- Term Peak Load Forecasting. IET Smart Grid, Under Review