Experimental and Machine Learning-Based Performance Prediction of TiN Nanofluid Heat Pipe Solar Collectors Under Indoor Test Conditions

Abstract

Solar thermal collector performance prediction plays an important role in enhancing system efficiency and developing sustainable energy systems. This paper develops a comprehensive machine learning-based framework for predicting the energy and exergy efficiency of a TiN-water nanofluid charged heat pipe-assisted solar thermal collector. Detailed experimental research was conducted in initial phase to generate a huge dataset containing approximately 500 steady-state operational configurations. This experimentally acquired information was used to train, validate, and test predictive models, laying the groundwork for subsequent machine learning analysis. Before developing the model, Boruta feature selection based on Random Forest significance was used to evaluate the relative relevance of operating parameters while maintaining all physically relevant inputs. Six supervised regression algorithms were implemented using a unified scikit-learn pipeline that included data standardisation and methodical hyperparameter optimisation via GridSearchCV: Multi-Layer Perceptron, Decision Tree, Random Forest, Gradient Boosting, K-Nearest Neighbours, and Extreme Gradient Boosting (XGBoost). Several statistical measures were employed to evaluate the model’s performance, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error, and mean absolute percentage error. The results clearly reveal that tree-based ensemble models outperform neural network and instance-based techniques. The Gradient Boosting model was the best accurate predictor of energy efficiency, with an R² of 0.9552 and an RMSE of 0.6592. XGBoost performed better at predicting exergy efficiency, with a R² of 0.9583 and an RMSE of 0.2609. When both performance goals were considered simultaneously, XGBoost was proven to be the most stable and reliable common model. Furthermore, heat flux and filling ratio were identified as the key regulating parameters utilising SHAP-based explainable artificial intelligence analysis, which is highly congruent with known thermodynamic principles and experimental data.