Optimised random forest for predicting bed expansion and pressure drop in media filter backwashing

Abstract

In drip irrigation systems, media filters are frequently used to avoid emitter clogging, but periodic backwashing is required to restore media filtration capacity. Backwashing reverses the flow, fluidises the media, and expels trapped particles, being bed expansion (BE) and pressure drop (PD) the key parameters for assessing backwashing hydraulic performance. Available equations and models, however, often fail to reliably predict these parameters under diverse operational conditions. This study introduces a machine learning-based model utilising Random Forest Regression (RFR) and the Whale Optimisation Algorithm (WOA) to predict both BE and PD from a dataset of 705 backwashing runs carried out in the laboratory. For comparison, Lasso, Elastic-net, and Ridge regression models were also implemented with the WOA optimiser. The RFR model was tuned to enhance accuracy by identifying key operational inputs: filter medium type (three categories: silica sand 0.75-0.85 mm, microspheres 0.63-0.75 mm, and silica sand 0.63-0.75 mm), underdrain type (four categories), filter bed height (200 or 300 mm), and superficial velocity. Results showed the WOA/RFR model not only ranked input variable importance but also achieved superior predictive accuracy, with coefficients of determination of 0.9771 and 0.9957 for BE and PD, respectively. The WOA/RFR model consistently outperformed Lasso, Elastic-net, and Ridge models, but demonstrated robust alignment with experimental data. This study presents a novel and optimised approach for predicting bed expansion and pressure drop, enhancing the reliability of media filter backwashing performance assessments in irrigation systems

Computational assistance was supplied by the University of Oviedo's Mathematics Department, while grants from the University of Girona (Spain) and the Spanish Research Agency, and the European Regional Development Fund, Belgium, respectively, were awarded PONT2022/03 and RTI2018-094798-B-100. Jonathan Graciano-Uribe thanks the IFUdG342022 fund for its support