Long-term time series forecasting with recurrent neural networks

Description

Neural networks are powerful function approximators that can capture and represent linear and nonlinear variations in your data. Recurrent neural networks (RNN) extend this approach by taking into account the sequential nature of the data. This sequential relationship is captured by creating an internal description of the input, which is computed by using not only the current observation in a series but also the current representation of all the points that have been observed so far.

Mustafa Kabul shares an approach for training an RNN on time series data that exhibits short- and long-term seasonality to generate multiple time steps ahead forecasts and explains how to transform a long time series input into shorter sequences that retain the patterns relevant for capturing the multiple modes of seasonality and the variability. This approach allows you to multithread and distribute the training process.

Mustafa demonstrates the long-term forecast performance using SAS’s solar farm production data and shows ways to incorporate derived features to enhance the time series for better accuracy, presenting results that generate much better long-term forecast accuracy than the traditional time series forecasting methods. This technique is extensible to practical applications like forecasting consumption demand and predicting stock markets and weather conditions.