Development of a digital twin in the absence of operational data: synthetic data generation and energy optimization of an industrial refrigeration system

This thesis investigates the development of a process-based digital twin for an industrial refrigeration system in the absence of operational plant data. While most digital twin frameworks rely on extensive instrumentation and historical datasets, many existing industrial facilities, particularly legacy and utility systems, lack such infrastructure entirely. This work addresses this gap by developing a physically consistent process model built exclusively from design documentation and project specifications, without any measured plant data. The proposed architecture comprises four elements: a steady-state hydraulic model of a ten-pump parallel system implemented in Python, an energy optimization framework based on the CMA-ES algorithm, a synthetic plant emulator introducing realistic model–plant mismatch through pump degradation and hydraulic fouling, and a neural-network-based adaptation module that learns discrepancy patterns between the ideal model and the emulated plant response. Applied to an annual cooling demand profile, the optimization achieved an approximate 27% reduction in pumping energy consumption compared to the current manual operating strategy. The adaptation module reduced the mean absolute flow rate error from 41.1 m³/h to 7.0 m³/h, demonstrating the ability to track equipment-specific degradation and restore adequate system performance. Additionally, the same case study was modeled in gPROMS Process (Siemens) to evaluate whether commercial simulation platforms can serve as plug-and-play tools for digital twin development in data-poor contexts. The analysis showed that standard simulation modes require operational information that is inherently unavailable in such environments, necessitating substantial auxiliary modeling effort that partially negates the ease-of-use advantage of a commercial platform. The results demonstrate that a functioning digital twin architecture can be established from limited design information alone, provided that the development is grounded in physically consistent modeling and supported by a self-updating adaptation mechanism.