A novel comparative study of machine learning surrogate models for solar chimney (SC) plant performance evaluation: Thermo-physical insights

This study analyzes the thermophysical performance of a conventional Manzanares solar chimney (SC) plant for finding the optimal configuration by varying three key design parameters, namely inlet height of the collector, diameter, and divergence of the chimney at different irradiation levels (400–1000 W/m2). A numerical solver based on finite volume methods is employed to run simulations. Additionally, multiple machine learning surrogate models were evaluated to identify the most effective approach for performance prediction. The analyses of three geometric parameters leads to optimum design values, which vary with solar irradiation. For a solar intensity of 1000 W/m2, the most efficient collector inlet height is about 0.2 m providing a ∼116 % power increase compared to the standard Manzanares plant. The optimal inlet height increases to 1.0 m at lower irradiation (400 W/m2). It is determined that under any irradiation conditions, chimney diameter increase beyond ∼45 m leads to a negligible improvement in power generation. This power generation is ∼318 % more compared to the Manzanares plant at 1000 W/m2 at this chimney diameter. The optimal outlet chimney diameter is approximately 15 m, which maximizes power generation for the Manzanares model at all incident solar radiation levels, resulting in a ∼52 % enhancement in performance compared to the standard Manzanares plant of straight chimney. Additionally, a comparative analysis of machine learning (ML) models, including decision trees, linear regression, artificial neural networks (ANN), support vector machines (SVM), and Gaussian process regression (GPR), demonstrates the superior predictive accuracy and robustness of GPR models. The iterative evaluation of GPR models using a Monte Carlo cross-validation approach confirms their reliability, with 25 iterations resulting in a mean R2 magnitude of 0.9 and 95 % lower RMSE values compared to other ML techniques, regardless of the chimney responses considered.