xadrianzetx / temporal-conv-net-keras

TCN architecture for time series

timeseries

TCN architecture for time series

Evaluation

Temporal convolutional network was evaluated with respect to following points:

• Split selected dataset into training, evaluation and testing subsets
• Test set has lenght of seq_length + n_ahead where seq_length is length of sequence fed to TCN, and n_ahead is prediction length
• Train network using training and evaluation sets
• Feed seq_lenght of test set into trained network and get n_ahead predictions
• Evaluate n_ahead predictions against n_ahead from test set

Evaluation set 1 - Historical Hourly Weather Data

Source

Evaluation set 2 - Air Quality in Madrid

Source

Evaluation set 3 - Hourly Energy Consumption

Source

TODO

• compare TCN to other models (LSTM, Holt Winters)

References

• Bai et al., 2018, An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling