Review on deep learning models for time series forecasting in industry

This review outlined recent developments in deep learning time series forecasting technology for the needs of industrial applications. With the advances in industrial automation, the storage and analysis of massive production data have become possible. Traditional mechanism modeling methods based on statistics encounter difficulties in dealing with high-dimensional industrial problems. Thus, time series forecasting for complex processes and products has played an important role in industrial scenarios, such as device modeling, production forecasting, remaining life prediction, and precise control, thereby receiving considerable attention from both academia and industry. To methodically review the time series forecasting method and its industrial applications, this review first introduced the three types of time series forecasting, namely, statistical learning, integrated learning, and deep learning, and compared their ease of use, complexity, and applicability issues. Focusing on industrial data analysis and decision-making, this review analyzed the three types of deep learning models, namely, recurrent neural network, convolutional neural network, and encoder–decoder network. The advantages and disadvantages of these three types of models and the applicable industrial environment were given, and how they can be embedded in industrial application sites to reduce costs and improve production efficiency was described. Afterward, to evaluate the performance of different algorithms clearly and comprehensively, statistical metrics and loss functions for the point prediction sequence and shape (motif) prediction problems were presented. At the same time, this review compiled classic public datasets of the industry for researchers to quickly find authoritative assessment data for an industry sector or issue. Taking mining and metallurgy in the process industry as examples, this review presented some widespread problems in this field, such as nonlinearity, long time delays, and unobservability of variables. This review also showed how deep-learning-based time series forecasting techniques can solve the aforementioned problems, build soft sensors, create digital twins, and achieve the visualization of complex processes. This review revealed that the application of deep learning in the process industry requires highly robust and strongly interpretable or explainable algorithms. For the robustness problem, the use of the ordinary differential equation model and Kalman filter method to solve the modeling of irregularly sampled time series and the use of the deep learning method to detect online sensor abnormalities were proposed. For the interpretability problem, sample-based, structure-based interpretable, and external co-explanation methods were introduced. This review also analyzed how explainable techniques can be applied to industrial deep learning models. Finally, the future directions of time series research were discussed in terms of both deep learning methods and industrial applications.