With the advent of Big Data, nowadays in many applications databasescontaining large quantities of similar time series are available. Forecastingtime series in these domains with traditional univariate forecasting proceduresleaves great potentials for producing accurate forecasts untapped. Recurrentneural networks, and in particular Long Short-Term Memory (LSTM) networks haveproven recently that they are able to outperform state-of-the-art univariatetime series forecasting methods in this context, when trained across allavailable time series. However, if the time series database is heterogeneousaccuracy may degenerate, so that on the way towards fully automatic forecastingmethods in this space, a notion of similarity between the time series needs tobe built into the methods. To this end, we present a prediction model usingLSTMs on subgroups of similar time series, which are identified by time seriesclustering techniques. The proposed methodology is able to consistentlyoutperform the baseline LSTM model, and it achieves competitive results onbenchmarking datasets, in particular outperforming all other methods on theCIF2016 dataset.
展开▼
