samsipe / forecastNet

Code for the paper entitled "ForecastNet: A Time-Variant Deep Feed-Forward Neural Network Architecture for Multi-Step-Ahead Time-Series Forecasting"

ForecastNet

Implementation of ForecastNet described in the paper entitled "ForecastNet: A Time-Variant Deep Feed-Forward Neural Network Architecture for Multi-Step-Ahead Time-Series Forecasting" by Joel Janek Dabrowski, YiFan Zhang, and Ashfaqur Rahman.

Link to the published paper: https://link.springer.com/chapter/10.1007/978-3-030-63836-8_48

Link to the Arxiv paper (older version but more detail): https://arxiv.org/abs/2002.04155

ForecastNet is a deep feed-forward neural network multi-step-ahead forecasting of time-series data. The model is designed for (but is not limited to) seasonal time-series data. It comprises a set of outputs which are interleaved between a series of "cells" (a term borrowed from RNN literature). Each cell is a feed-forward neural network which can be chosen according to your needs. This code presents ForecastNet with two different cell architectures: one comprising densely connected layers, and one comprising a convolutional neural network (CNN).

The key benefits of ForecastNet are:

1. It is a time-variant model, as opposed to a time-invariant model (In the paper we show that RNN and CNN models are time-invariant).
2. It naturally increases in complexity with increasing forecast reach.
3. It's interleaved outputs assist with convergence and mitigating vanishing-gradient problems.
4. The "cell" architecture is highly flexible.
5. It is shown to out-perform state of the art deep learning models and statistical models.

Usage Notes

• A PyTorch implementation and a TensorFlow implementation are provided. The PyTorch implementation is recommended as it is is more complete and more generic.
• Both implementations provide a demonstration using a synthetic dataset. The PyTorch implementation will be easiest to adapt to your own dataset.
• The TensorFlow implementation is written for univariate time-series. The PyTorch implementation accepts multivariate datasets (it has been tested with univariate datasets and datasets with multivariate inputs and a univariate output).
• Please read the README files in each implementation's directory.

Citation

Dabrowski J.J., Zhang Y., Rahman A. (2020) ForecastNet: A Time-Variant Deep Feed-Forward Neural Network Architecture for Multi-step-Ahead Time-Series Forecasting. In: Yang H., Pasupa K., Leung A.CS., Kwok J.T., Chan J.H., King I. (eds) Neural Information Processing. ICONIP 2020. Lecture Notes in Computer Science, vol 12534. Springer, Cham. https://doi.org/10.1007/978-3-030-63836-8_48

@InProceedings{10.1007/978-3-030-63836-8_48,
    author="Dabrowski, Joel Janek
    and Zhang, YiFan
    and Rahman, Ashfaqur",
    editor="Yang, Haiqin
    and Pasupa, Kitsuchart
    and Leung, Andrew Chi-Sing
    and Kwok, James T.
    and Chan, Jonathan H.
    and King, Irwin",
    title="ForecastNet: A Time-Variant Deep Feed-Forward Neural Network Architecture for Multi-step-Ahead Time-Series Forecasting",
    booktitle="Neural Information Processing",
    year="2020",
    publisher="Springer International Publishing",
    address="Cham",
    pages="579--591",
    abstract="Recurrent and convolutional neural networks are the most common architectures used for time-series forecasting in deep learning literature. Owing to parameter sharing and repeating architecture, these models are time-invariant (shift-invariant in the spatial domain). We demonstrate how time-invariance in such models can reduce the capacity to model time-varying dynamics in the data. We propose ForecastNet which uses a deep feed-forward architecture and interleaved outputs to provide a time-variant model. ForecastNet is demonstrated to model time varying dynamics in data and outperform statistical and deep learning benchmark models on several seasonal time-series datasets.",
    isbn="978-3-030-63836-8"
}