Time Series Representation Learning: A Survey on Deep Learning Techniques for Time Series Forecasting

With the rise of Industrial Internet of Things (IIoT) and other aspects of the digitalization of production more and more time series (TS) data is generated. This data can help to train deep learning models with higher cognitive capabilities and thus assist with more complicated tasks. An important step for creating such models is creating fitting abstract representations of the data. Thus, this work surveys the literature on deep learning techniques for time series forecasting. The focus hereby lies on the characteristics and reasons for using the techniques in context of the representation learning. 17 of the architectures used in the analyzed literature use multiple techniques. The most applied techniques are one-dimensional CNN (CNN 1D) (14 times), LSTM (11 times), and attention-based techniques (17 times). Furthermore, input embedding and masking play an important role in some architectures.