From Prediction to Interaction: Exploring Deep Learning Forecasting Models for Human-Centered Industrial Applications

The integration of Artificial Intelligence (AI), particularly Deep Learning (DL), is driving a paradigm shift in industrial practices, moving beyond traditional automation and production management. In the era of Industry 5.0, the focus is shifting toward human-centered systems that not only improve efficiency but also promote safety, adaptability, and sustainability. Deep Learning plays a key role in this transition, offering the ability to process complex, multimodal data and provide real-time insights in line with the dynamic nature of modern industrial environments. By enabling systems to anticipate, adapt, and interact, Deep Learning transforms prediction from a passive tool to an active tool of Human Robot Collaboration (HRC). Traditional forecasting methods, while effective in structured and predictable scenarios, often struggle with the complexity and variability inherent in dynamic industrial applications. This thesis investigates how Deep Learning models can bridge this gap by progressively scaling capabilities across different applications of Time-Series Forecasting: from production planning to adaptive systems, and ultimately to real-time interaction between robots and humans. The contributions of this thesis are demonstrated through three key industrial tasks, each exemplifying the main challenges of the design of a more sustainable, safe, and human-centered industry. The first tackled task is called New Fashion Product Performance Forecasting (NFPPF), addressing the complexity of predicting demand for garments with no historical sales data. This task highlights how can support sustainable production planning, reduce waste, and align supply with market needs. The second task explores human trajectory forecasting in dynamic indoor environments, such as warehouses and factories, where understanding and predicting human motion is essential for operational safety and efficiency. By enabling systems to adapt to unpredictable movements, this task lays the foundation for safer Human Robot Collaboration. Finally, this thesis investigates Human Pose Forecasting (HPF), a critical element in enabling seamless interaction between humans and robots. In scenarios such as assembly lines or shared workspaces, accurate predictions of human movement in real-time allow robots to respond proactively, ensuring safety and enhancing cooperative workflows. Lastly, this thesis highlights the potential of Deep Learning in aligning technological advances with human-centered principles, offering a roadmap for industries to embrace the transformative vision of Industry 5.0 where Artificial Intelligence and humans work together to redefine the future of work and manufacturing.