Prediction of tubular solar still performance by machine learning integrated with Bayesian optimization algorithm

• Two machine learning models were developed to predict the productivity of tubular solar still. • Bayesian optimization algorism was considered for both models. • Optimized models more accurately predicted production with better evaluation indicators. • Random forest was less sensitive to hyper parameters compared to artificial neural network. In this study, accurate and convenient prediction models of tubular solar still performance, expressed as hourly production, were developed by utilizing machine learning. Based on experimental data, the models were developed and compared, such as classical artificial neural network with/without Baysian optimization, random forest with/without Baysian optimization, and traditional multilinear regression. Before applying Bayesian optimization, both random forest and artificial neural network predict hourly production. But the superiority of random forest is well behaved with insignificant error. The prediction performance of random forest, artificial neural network and multilinear regression were calculated as 0.9758, 0.9614, 0.9267 for determination coefficients, and 5.21%, 7.697%, 10.911% for mean absolute percentage error, respectively. Additionally, when applying Bayesian optimization for searching most appropriate hyper parameters, the performance of artificial neural network was significantly improved by 35%. Moreover, optimization findings revealed that random forest was less sensitive to hyper parameters than artificial neural network. Based on the robustness performance and high accuracy, the random forest is recommended in predicting production of tubular solar still.