Data-driven prediction of hemispherical solar distiller performance: Optimizing water production with machine learning

Abstract

Artificial intelligence has found applications in numerous research domains, notably in the field of mechanical engineering. This research work introduces a novel data-driven approach to predict the thermal performance of a Hemispherical Solar Still (HSS) using five machine learning models: Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN). The proposed prediction models are built using real experimental data that has been recorded. Leveraging real experimental data, these models were rigorously evaluated for predicting hourly productivity and instantaneous efficiency, using five statistical error metrics to assess their performance. The DT model emerged as the most accurate and efficient, achieving near-ideal R² and EVS values near to one, alongside minimal statistical error values (MSE, NAE, and Median Absolute Error). Notably, DT predicted average hourly productivity at 0.477 L/m²/day—closely matching experimental averages (0.478 L/m²)—with 45.2 % average instantaneous efficiency, outperforming other models (SVM: 46.4 %, RF: 46.4 %, KNN: 44.8 %, GB: 45.2 %). Therefore, this work demonstrates that DT-based prediction can reliably estimate HSS performance, eliminating costly and time-consuming experimental iterations. The proposed framework provides a robust, scalable solution for solar still optimization, advancing AI applications in sustainable water production.