A new predictive model for the outlet turbidity in micro-irrigation sand filters fed with effluents using Gaussian process regression

• Published in: Computers and electronics in agriculture 2020-03, Vol.170, p.105292, Article 105292
• Main Authors: García Nieto, P.J., García-Gonzalo, E., Puig-Bargués, J., Solé-Torres, C., Duran-Ros, M., Arbat, G.
• Format: Article
• Language: English
• Subjects: Bayesian statistics; Clogging; Drip irrigation; Electrical resistivity; Emitters; Gaussian process; Gaussian process regression (GPR); Irrigation; Machine learning techniques; Optimization; Optimization techniques; Performance prediction; Prediction models; Process parameters; Regression analysis; Regression models; Sand; Sand filters; Turbidity; Water temperature
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105292

•Prediction of sand filter outlet turbidity values allows assessing the reusing effluents in irrigation.•A hybrid model based on GPR with the LBFGSB optimization technique was used for this prediction.•The developed model predicted satisfactorily the sand filter outlet turbidity.•The obtained correlation coefficient from this hybrid GPR–based model is about 90%. Sand media filters used in microirrigation systems must remove suspended particle load for avoiding emitter physical clogging. Turbidity is a parameter related to suspended particle load that it is easy and quick to measure and it is also included in some guidelines for reusing effluents in irrigation. Currently, there are not sufficiently accurate models available to predict outlet turbidity for sand filters, which would be useful for both irrigators and engineers. The aim of this study was to obtain a predictive model able to perform an early detection of the sand filter outlet value of turbidity. This study presents a powerful and effective Bayesian nonparametric approach, termed Gaussian process regression (GPR) model, for predicting the output turbidity (Turbo) from data corresponding to 637 samples of different sand filters using reclaimed effluent. This optimization technique involves kernel parameter setting in the GPR training procedure, which significantly influences the regression accuracy. To this end, the most important parameters of this process are monitored and analyzed: type of filter, height of the filter bed (H), filtration velocity (v) and filter inlet values of the electrical conductivity (CEi), dissolved oxygen (DOi), pHi, turbidity (Turbi) and water temperature (Ti). The results of the present study are two-fold. In the first place, the significance of each variable on the filtration is presented through the model. Secondly, a model for forecasting the outlet turbidity was obtained with success. Indeed, regression with optimal hyperparameters was performed and a coefficient of determination equal to 0.8921 for outlet turbidity was obtained when this new predictive GPR–based model was applied to the experimental dataset. The agreement between experimental data and the model confirmed the good performance of the latter.