Deep Learning-Based Time Series Forecasting for Industrial Discrete Process Data

With the introduction of Industry 4.0, continuous collection and monitoring of industrial process data have become fundamental aspects of modern operational technology (OT) systems. The ability to acquire high-resolution multivariate time series data offers new opportunities for data-driven forecasting. Such forecasting facilitates proactive interventions, including process optimization and anomaly detection, with applications ranging from efficiency improvements to cybersecurity. While deep learning-based forecasting models have demonstrated strong performance in various domains, their effectiveness for discrete manufacturing processes are not yet well studied. In this paper, we summarize key characteristics of discrete manufacturing process data and introduce a publicly available dataset that captures these key characteristics. Consequently, we evaluate six state-of-the-art deep learning-based forecasting models on this dataset. Notably, the performance ranking among the models is different than on typical benchmark datasets, reflecting the distinct characteristics of the introduced dataset. We conclude that Crossformer and DUET are particularly effective at generating short-term time series forecasts with low error. However, for applications requiring long-term forecasting via recursive prediction under real-time constraints, forecast accuracy rapidly decreases. To mitigate this, we introduce a second training pass that includes the model's own forecasts, which significantly reduces the error. Despite this, long-term forecasts remain challenging and further research is needed.