Session: 09-17-01: AI for Energy

Paper Number: 167746

Hybrid Approach Based on Simulation and Experimental Data for Surrogate-Based Modeling: A Case Study of a High-Temperature Heat Pump 

Surrogate models serve as approximations of complex, high-order simulation models, capturing their input-output behavior with significantly reduced computational effort. These models are particularly useful in various applications such as design and operational optimization, setpoint optimization (energy management), and predictive control, where repeated evaluations of detailed simulations would be expensive. In recent years, the role of surrogate models has grown considerably, as industries and researchers seek efficient methodologies to describe system components or entire processes in a computationally feasible manner. However, surrogate models based exclusively on simulation data often suffer from inaccuracies and uncertainties. These limitations arise due to two main factors: First, simulation models may not capture all relevant effects that occur in real-world systems. Second, deviations between simulation results and real process data are inevitable, leading to potential discrepancies between model predictions and actual performance. On the other hand, relying solely on experimental data presents its own challenges: collecting sufficient experimental data is often expensive and time-consuming, while measurement noise and inconsistencies can introduce additional uncertainties. To address these challenges, the integration of experimental data with simulation-based surrogate models has emerged as a promising approach. By leveraging real-world measurements, hybrid modeling techniques enhance the accuracy and reliability of surrogate models, mitigating the risks associated with purely simulation- and experimental-based methods. Hybrid modeling systematically integrates both simulation-generated and experimentally obtained data, ensuring that the surrogate model accurately represents the real system behavior while maintaining computational efficiency. The present study focuses on the development of a hybrid modeling strategy using statistical and AI methods to construct highly accurate surrogate models. Specifically, we investigate several methods, including transfer learning and adaptive weighting schemes that adjust based on the proportion of simulation and experimental data, as well as the level of data noise. To evaluate the effectiveness of these methods, we compare multiple hybrid modeling techniques in terms of accuracy and CPU training time. Specifically, we test these methods on analytical test cases with different levels of nonlinearity and different dimensions, as well as on a real-world use case of a high-temperature heat pump. Additionally, we assess their performance under different noise levels and varying proportions of simulation and experimental data. Through this comparative analysis, we explore the most effective strategy for building reliable and robust surrogate models. The results of this study demonstrate that the hybrid modeling approach significantly improves the predictive accuracy compared to surrogate models based solely on simulation or experimental data. By effectively combining both data sources, the hybrid model achieves improved generalization, lower uncertainty, and greater robustness. These findings highlight the potential of hybrid modeling strategies to enhance the reliability of surrogate models, making them more suitable for real-world applications in engineering and industrial optimization.

Presenting Author: Loukas Kyriakidis German Aerospace Center (DLR)

Presenting Author Biography: Loukas Kyriakidis completed his Bachelor’s and Master’s degree in Mechanical Engineering at the Technical University of Berlin in 2017 and 2019, respectively, and has worked as a research associate at the German Aerospace Center, at the Institute of Low-Carbon Industrial Processes since 2021. His research primarily focuses on the optimal control of utility systems, which includes optimization methods, surrogate modeling, and forecasting of input data through machine learning.

Authors:

Loukas Kyriakidis German Aerospace Center (DLR)
Ashish Thapa German Aerospace Center (DLR)