Predictive supervised machine learning models for double-slope solar stills

There has been a growing trend to develop predictive solar-desalination models via machine learning (ML) and artificial intelligence (AI) tools. However, forecasting productivities of solar stills of different designs remains a challenge that can be overcome by establishing regression correlations via built-in and pre-existing ML toolboxes. Herein, the author developed accurate supervised predictive ML models for the productivity predictions in a double-slope still based on literature experimental results. Training datasets were constructed from the earlier observations (inputs/outputs) from various designed passive and/or active solar stills which were used to treat brackish water or wastewater with 45% TDS. A semi-proportional relationship between water-glass temperature (Tw - Tg) and water distillate was established with a minimum statistical error. The relationship proposed that an increase in both Tw - Tg and basin temperature (TB) would result in the maximum distillate at time 14:00. The regression models (FGSVM, EBoT, SEGPR) showed that they had the least achieved root mean square error (RMSE) of <138 indicating their reliability to accurately predict the distillate amounts in double-slope designs. The high accuracy of the SEGPR trained model with (R2 = 1) and the very low RMSE < 8 showed the capability of the model to predict the performance in similar solar-desalination systems. Yet, the FGSVM was found to be more reliable in predicting Tw - Tg whereas that the stepwise linear regression (SLR) better predicted the TB pattern against the water distillate. This work suggests the importance of both the FGSVM and the SLR models for water outputs predictions which can pave the way towards establishing a unified theoretical tuning-parameters model to maximize the performance and the distillate water in double-slope solar stills.