Development of a robust data-driven surrogate model to improve energy flexibility of an integrated district heating system with a thermal storage system

Data-driven surrogate models (SMs) are increasingly applied to optimize energy systems, yet the impact of dataset variability and diversity on model robustness is often overlooked in operational optimization. Phase change material (PCM) tanks can enhance energy flexibility of district heating (DH) systems, but an integrated strategy that considers both spot price fluctuations and peak demand remains underexplored. To bridge this research gap, this study presents a robust SM to enhance the performance of a DH system integrated with a data centre and a PCM thermal storage system. Various operational strategies for PCM tank charging and discharging were implemented to generate diverse datasets for training the SMs and identifying the most robust model. Long short-term memory networks were used to develop the models, following careful input-output selection. A new operational strategy was then proposed, focusing on peak shaving and load shifting, using the best-performing SM to optimize tank operation. A case study showed that the day-ahead price-based control strategy improved model robustness by enabling dynamic day-to-day adjustments in charging and discharging, leading to more effective management and the generation of well-distributed datasets. The tailored strategy achieved an 11.5 % reduction in peak demand, a 0.3 % decrease in price-responsive heating costs, and a 6.1 % reduction in total operational costs compared with the daily time-based strategy. This study provides a reference for developing robust SMs using datasets generated by demand response strategies and offers guidance on improving the performance of hybrid DH systems.