Early stopping technique using a genetic algorithm for calibration of an urban runoff model

Identifying suitable parameter sets for use in catchment modelling remains a critical issue in hydrology. This paper describes an early stopping technique (EST) for use during calibration of a multi-parameter urban catchment modelling system. The proposed method takes advantage of MODE and lower confidence limit (LCL) functions in statistical analysis of spanning set of objective function values. The paper also introduces a monitoring process and regularization techniques to avoid under/overfitting during the calibration and to enhance generalisation performance. The methodology is assessed using SWMM and linked with a Genetic Algorithm for calibration of a Powells Creek catchment model in Sydney, Australia. Results demonstrate that the statistical spanning set analysis approach overcomes issues of poor interpretation and deterioration in the model’s generalisation properties. By stopping early, the calibration process avoided overfitting; this was indicated by too closely fitting to the calibration dataset and a failure to fit to the monitoring dataset.