Robust Time Series Prediction with Missing Data Based on Deep Convolutional Neural Networks

Abstract

Recurrent Neural Network (RNN) is a class of neural networks for processing sequential data. Accordingly, when we predict some long-term sequence information, such as the flight information within one year, we usually use the RNN. However, as a general recurrent neural network, RNN cannot deal with the temporal data involving a mixture of long-term and short-term patterns. Therefore, we adopt a new deep learning framework, that is, long- and short-term time series network (LSTNet), which is composed of CNN, RNN, skip RNN and AR component. On this basis, the existing LSTNet algorithm only uses the observed data to predict time series without considering the robustness of the overall structure. The prediction error will be greatly increased when some of the data is missing. In this paper, we propose a novel deep learning framework based on LSTNet network, namely Missing data LSTNet network (M-LSTNet), to solve the problem of time series prediction in the presence of missing data in LSTNet network. Compared with the original framework, we add two new algorithms, M-Impute and M-ARIMA. The algorithm M-Impute is used to judge whether the missing data has occurred and compensate the discontinuous time series containing missing data into continuous time series. The later algorithm M-ARIMA uses the time series predicted by ARIMA to replace the continuous time series obtained in the M-Impute so as to improve the original LSTNet framework and solve the negative impact caused by data missing. Based on the deep learning framework M-LSTNet in this paper, we calculate the prediction evaluation of the original time series, the time series containing missing data and the time series improved by our algorithms. The results show that, our compensation algorithm can obtain better prediction effect and improve the stability of the whole deep learning network.