The problem of predicting nonlinear and nonstationary signals is complex since the physical law that controls them is unknown and it is complicated to be considered. In these cases, it is necessary to devise nonlinear models that imitate or learn the rules of behavior of the problem and can be developed based on historical data. For this reason, neural networks are useful tools to deal with this type of problem due to their nonlinearly and their capacity of generalizing. This paper aims at exploring various types of neural network architectures and study their performance with time series predictions. Predictions on two sets of data (of a very different nature) will be made using three neural networks including multilayer perceptrons, recurrent neural network and long-short term memory varying some important parameters: input neurons, epochs and the anticipation with which the predictions are made. Then, all results will be compared using standard metrics. As a conclusion, the influence of the type of series under study is more important than the parameters considered in what concerns the performance. The management of the memory in the networks is a key to its success in the prediction of S&P 500 and electrical power time series.
|Title of host publication||Intelligent Computing Theories and Application - 15th International Conference, ICIC 2019, Proceedings|
|Editors||De-Shuang Huang, Vitoantonio Bevilacqua, Prashan Premaratne|
|Number of pages||11|
|Publication status||Published - 24 Jul 2019|
|Event||15th International Conference on Intelligent Computing, ICIC 2019 - Nanchang, China|
Duration: 3 Aug 2019 → 6 Aug 2019
|Name||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Conference||15th International Conference on Intelligent Computing, ICIC 2019|
|Period||3/08/19 → 6/08/19|
Bibliographical notePublisher Copyright:
© 2019, Springer Nature Switzerland AG.
- Long short-term memory
- Neural Turing Machine