Machine learning models for performance estimation of solar still in a humid sub-tropical region
摘要
In the current investigation, machine learning models were developed to estimate the performance of a solar still in a humid subtropical climate region. A single-slope passive solar still was designed and constructed to facilitate year-round experiments and data recording. The output variable under consideration was the Hourly Yield, with four parameters serving as input variables: global solar radiation, water glass temperature difference, ambient temperature, and wind speed. 485 h of data were employed in the machine learning phase to assess the solar still’s performance. Five machine learning models, namely, k-nearest neighbors, support vector machines, random forest, multilayer perceptron, and extreme gradient boosting, were employed to predict hourly yield. Their performance was evaluated using statistical indicators such as mean bias error, mean absolute percentage error, correlation coefficient, t-statistics, and maximum absolute relative error. The results showed that random forest was the most reliable model, achieving the highest correlation coefficient (0.9812) and the lowest mean absolute percentage error (8.8221 MJ/m2 day). Extreme gradient boosting followed closely with similarly high accuracy, while k-nearest neighbors performed the weakest, with the lowest correlation coefficient (0.8992) and the highest mean absolute percentage error (20.34%). Statistical metrics were normalized to compute the global performance indicator to further assess model performance. The multilayer perceptron and random forest models emerged as the top performers, with global performance indicator scores of 0.801 and 0.778, respectively. The validation of these models was conducted using experimental data.