Forecast of the outlet turbidity and filtered volume in different microirrigation filters and filtration media by using machine learning techniques

Abstract

Different media filters and filtration media are used to eliminate suspended particles in microirrigation and therefore prevent clogging in the emitter. The water volume by filtration cycle is a parameter related to the filter and media capacity to retain particles, while turbidity is a variable related to particles in suspension in the water. Since turbidity can be measured easily and quickly, it is commonly mentioned in recommendations for the reuse of effluents in microirrigation. There are currently no models that are reliable enough to predict the filtered volume in each irrigation filter and outlet turbidity when using different filter types and media configurations. The object of this work was to propose a model that can detect early the filtered volume and the turbidity at the outlet values. This investigation presents an effective machine learning method, the Random Forest regression (RFR) in combination with the population-inspired metaheuristic optimization algorithm, called Differential Evolution (DE), for estimating the output turbidity and the filtered volume from a dataset with 1,016 samples of distinct media filters that use reclaimed effluent. The same experimental dataset was also fitted with Elastic-net, Lasso and Ridge regression machine learning methods also in combination with DE optimizer for comparison. This optimization performs the parameter tuning in the RFR using the training dataset, which considerably improves the accuracy of the regression. To achieve this, the most relevant operation input variables are tracked and analyzed: the kind of medium and filter, filtration velocity (v), height of the filter bed (H), cycle duration and the electrical conductivity for the filter inlet (EC ), pH, dissolved oxygen (DO), water temperature (T) and the input turbidity (Turb). There are two kinds of results. Firstly, the importance ranking of the input variables on the outlet turbidity and filtered volume is presented using the DE/RFR model. Secondly, an innovative model for the prediction of the outlet turbidity and filtered volume was built and a regression with optimized parameters was done and coefficients of determination of 0.9331 and 0.8712 for filtered volume and outlet turbidity were obtained with this DE/RFR-based model, respectively. Additionally, the outcomes from the Elastic-net, Lasso and Ridge models are worse than DE/RFR-relied model estimations. The DE/RFR-based model's strong performance was confirmed by the agreement between experimental data and the latter results