A Novel Prediction-Optimization Machine Learning Framework for Nanofluid-Based Photovoltaic/Thermal Systems

Nanofluid-based spectral filtering offers a promising approach to enhance photovoltaic/thermal (PV/T) system performance by utilizing the full solar spectrum. However, system optimization remains challenging due to complex nonlinear relationships between nanofluid parameters and overall performance. This study develops a prediction-optimization framework integrating deep neural networks (DNN) with genetic algorithms (GA) to accurately analyze multi-parameter interactions and achieve globally optimal designs for nanofluid-based PV/T systems. High-throughput datasets for three nanofluids (Ag, Au, Al) were constructed using theoretical calculations that combined Lorentz-Mie theory, Monte Carlo simulations, and a coupled opto-electro-thermal model. Three machine learning models-DNN, random forest (RF), and decision tree (DT)-were employed to predict key PV/T performance parameters. By synergizing machine learning with GA, a closed-loop prediction-optimization process was established to efficiently identify optimal design parameters. Among the models evaluated, the DNN demonstrated superior performance, achieving prediction accuracies above 99.48% for all three key performance indicators (η_pv, η_th, and MF), significantly outperforming the RF and DT models. Furthermore, SHAP analysis was conducted to quantify the contribution of each input feature and enhance model interpretability. Coupled with the GA, the DNN-GA framework successfully identified globally optimal design parameters for each nanofluid. For instance, for Ag nanofluid, the optimal combination (r = 4.02 nm, h = 9.91 mm, f_v = 9.45 × 10^-5) yielded a maximum MF value of 1.3603. This work presents an innovative machine learning framework for designing nanofluid filters in PV/T systems, which reduces reliance on iterative experimentation and accelerates the development of high-performance solar energy systems, demonstrating practical value.