A0731
Title: Predictive models for time series by deep structured learning
Authors: Shubin Wu - University of Southampton (United Kingdom) [presenting]
Zudi Lu - University of Southampton (United Kingdom)
Abstract: Empirical studies on time series forecasting with deep neural networks have become widespread, but there is still no clear conclusion on which method works well, posing significant mysteries. The impact of the structure of a deep learning model on prediction, specifically examining the asymptotic properties of a multiple-layer neural network parametric autoregression under time series beta-mixing conditions, is investigated. An hourly wind electricity production time series dataset is used to compare three popular models: autoregressive multiple layer perceptron deep learning (AR-MLP), long-short term memory neural network (LSTM), and residual neural network (ResNet). The accuracy of these models is measured in terms of mean square error (MSE) for the predictions, and an interpretable model that reasonably fits the hidden layers of the neural network is explored through a reasonable selection of the AR lag order for the input with an equal number of neurons in each hidden layer and the number of hidden layers. The results show that the AR-MLP model achieves high prediction accuracy and efficiency after considering validation set MSE, test MSE, and training time, suggesting that a neural network model with a relatively simple structure can potentially provide good prediction performance.