Feature and Architecture Selection on Deep Feedforward Network for Roll Motion Time Series Prediction

Abstract

The neural architecture and the input features are very substantial in order to build an artificial neural network (ANN) model that is able to perform a good prediction. The architecture is determined by several hyperparameters including the number of hidden layers, the number of nodes in each hidden layer, the series length, and the activation function. In this study, we present a method to perform feature selection and architecture selection of ANN model for time series prediction. Specifically, we explore a deep learning or deep neural network (DNN) model, called deep feedforward network, an ANN model with multiple hidden layers. We use two approaches for selecting the inputs, namely PACF based inputs and ARIMA based inputs. Three activation functions used are logistic sigmoid, tanh, and ReLU. The real dataset used is time series data called roll motion of a Floating Production Unit (FPU). Root mean squared error (RMSE) is used as the model selection criteria. The results show that the ARIMA based 3 hidden layers DNN model with ReLU function outperforms with remarkable prediction accuracy among other models.