Multivariate linear regression versus symbolic regression from genetic programming. Application to the spectroscopic characterisation of urban wastewater

Characterising urban wastewater in real time is key to ensure the proper management of water resources and environmental protection. From indirect measurements, such as the molecular spectroscopy which provides information on the physicochemical properties of the water, it is possible to determine the pollutant load of wastewater from mathematical correlation models. The research compares multivariate linear regression models and symbolic regression models based on genetic programming to establish a correlation with the pollutant load of the wastewater. The study has focused on the comparison of models for the characterisation of total nitrogen, total phosphorus and nitrogen in the form of nitrate of 90 urban wastewater samples. It is observed that the symbolic regression based on genetic programming provides an improvement in goodness of fit (R2) of between 72.76% and 146.39% with respect to multivariate linear regression.