Two Phase Multivariate to Multivariate Time Series Forecasting Using Self-attention Convolutional Autoencoder and Temporal Convolutional Network

Abstract

In manufacturing process, data is collected in the form of correlated sequences. Multivariate to multivariate time series (MMTS) forecasting is an important factor in manufacturing. MMTS forecasting is a notoriously challenging task considering the need for incorporating both non-linear correlations between variables (inter-relationships) and temporal relationships of each univariate time series (intra-relationships) while forecasting future time steps of each univariate time series (UTS) simultaneously. However, previous works use deep learning models suited for low-dimensional data. These models are insufficient to model high-dimensional relationships inherent in multivariate time series (MTS) data. Furthermore, these models are less productive and efficient as they focus on predicting a single target variable from multiple input variables. Thus, we proposed two phase MTS forecasting. First, the proposed method learns the non-linear correlations between UTS (inter-relationship) through self-attention based convolutional autoencoder and conducts cause analysis. Second, it learns the temporal relationships (intra-relationships) of MTS data through temporal convolutional network and forecasts multiple target outputs. As an end-to-end model, the proposed method is more efficient and derives excellent experimental results.