Rainfall–runoff modelling using genetic programming

This paper presents the application of genetic programming to the generation of models to assess the total runoff of a basin starting from the total rainfall in it and using data recorded in a sub-basin at the valley of Mexico (the Mixcoac sub-basin to the west of Mexico City). The modelling process is developed contrasting two types of models with different complexity degree: (1) a nonlinear model whose complexity is resolved using multi-objective optimization and (2) a nonlinear model with a given structure obtained by means of a physical interpretation of the dynamics of the direct and the base flow. Data from two storms (rainfall and runoff), one in 1997 and another in 1998, were used in testing the models. First, the storm in 1997 was used for the calibration step and that in 1998 for the validation step. Afterwards, the order was reversed. An interpretation of the results, focused on the applicability and possible improvement of the models in forecasting runoff, is made through their discussion and is summarized in the conclusions.