Constrained Gaussian processes as a surrogate model for simulation-based optimization of solar process heat systems

Assessing the performance of renewable energy powered systems, such as solar process heat, may require the development of complex and computationally demanding models capable of simulating such systems under numerous different design configurations and operating conditions. The resulting grid of data is often used to train a surrogate model representing the system's performance, which can be used to further analyze the system's profitability and calculate economically optimal configurations. This approach, however, may be unfeasible since it requires too many simulations. Therefore, this study proposes a novel and alternative approach, employing constrained Gaussian processes and active learning strategies to generate such surrogate models using significantly less training data. The results show that, for a 4D example case tested, the proposed method has generated reliable surrogates models using around 75 % less training data and requiring around half of the simulation time when compared with a grid-based approach, while presenting similar performance in terms of well-established figures of merit, such as RMSE and maximum error.