NBEATS model implementation in TF2/Keras for time series prediction

The only dependency is tensorflow 2.x and ususal way to build model is to import model class and create the object instance by passing the time series dataframe.

import pandas as pd
import tensorflow as tf
from tcnn import ModelNBEATS

train = pd.read_csv('train.csv')
train.index = pd.date_range(start='2010-01-01',periods=146, freq='D')

model = ModelNBEATS(y=train)            # use default values
model.fit()
model.predict(10)
...

###### you can access tf model object via model.model_
model.model_.summary()

model.save('save_path')
######
model.load('load_path')

model = ModelNBEATS(y=train, {'xreg': 0, 
                            'scale_data':True,
                            'num_layers':4,
                            'num_neurons':128,
                            'stacks':'generic, trend, seasonality, generic':,
                            'activation':'relu',
                            'optimizer':'adam',
                            'num_epoch':100,
                            'loss':'mse'
})

• xreg: pd.DataFrame. DataFrame with all external regressors (for prediction purposes, additional periods must be included)
• scale_data: bool, whether to use standard scaler to scale data to 0-1 range; default True
• num_layers: int, number of dense layers to use in NBEATS block; default 4
• num_neurons: int, number of neurons within each dense layer; default 128
• stacks : list or str, list or comma separated str of values for stacks; default: 'generic, trend, seasonality, generic'
• activation: str. The activation used in the residual blocks o = activation(x + F(x)).
• optimizer: str/keras. Optimizer to use when training the model, either str or keras optimizer object - default 'adam'.
• num_epoch: int. Number of epochs to train the model on - default 100
• loss: str/keras. Loss function to use when training the model - deault 'mse'

Pandas DataFrame with time dimension in the rows and unique time-series in columns. Row index sould be date/date-time based.

Pandas DataFrame. By calling model.predict(num_steps), pandas dataframe will be generated with num_steps rows and same number of columns as the initial training dataset.

git clone [email protected]:apantovic/nbeats-tf2.git && cd nbeats-tf2
virtualenv -p python3 venv
source venv/bin/activate
pip install tensorflow==2.5.0