<div class="csl-bib-body">
<div class="csl-entry">Fränkel, C. (2018). <i>Long short-term memory neural networks for one-step time series forecasting</i> [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2018.45860</div>
</div>
-
dc.identifier.uri
https://doi.org/10.34726/hss.2018.45860
-
dc.identifier.uri
http://hdl.handle.net/20.500.12708/1847
-
dc.description.abstract
This master thesis aims to employ Long Short-Term Memory (LSTM) neural networks for one-step and multi-step time series forecasts. For this endeavor, a software stack, including a deep learning framework is selected and different machine learning- and statistical models are implemented. The performance of the LSTM approaches is compared to carefully chosen benchmark methods on an exemplary real-world problem and the experiments are run on powerful, cloud-based machines. To provide a methodological framework for time series forecasting projects, a seven-phase process model is elaborated. Further, to allow for model selection of computationally intensive deep learning methods under limited resources, a modified form of blocked cross-validation, together with a multi-stage Bayesian hyperparameter optimization approach is proposed. The proof of concept of the pro-posed methodology is conducted on the real-world problem in the domain of electricity demand forecasting. The implemented LSTM model clearly outperformed the benchmark models on all per-formance measures in the one-step walk forward out-of-sample test and showed a roughly 10% low-er root-mean square error than the second-best model which utilized double seasonal Holt-Winters exponential smoothing. Inspired by work in the area of natural language processing, for the multi-step scenario, an encoder-decoder LSTM neural network was implemented, as simpler architectures showed disappointing results. Also, the multi-step LSTM forecaster proofed to be a competitive ap-proach, but the purely statistical model had the lead. However, due to resource constraints, it was not possible to retrieve statements on the same validity level as for the one-step case. By comparing the LSTM-forecaster to the predictive performance of simple recurrent neural networks, the added value of the more complex, gated cell architecture of the LSTM has been indicated. A downside of LSTM neural networks is the relatively long training time which can be a problem for exhaustive hy-perparameter searches. On the other hand, LSTM neural networks showed to have good generaliza-tion ability and needed comparably infrequent retraining.
en
dc.language
English
-
dc.language.iso
en
-
dc.rights.uri
http://rightsstatements.org/vocab/InC/1.0/
-
dc.subject
LSTM
en
dc.subject
forecasting
en
dc.subject
time series
en
dc.subject
bayesian hyperparameter optimization
en
dc.subject
deep learning
en
dc.subject
neural networks
en
dc.subject
Internet of Things
en
dc.title
Long short-term memory neural networks for one-step time series forecasting