Abstract
The complex topography and inherent nonlinearity affiliated with influential hydrological processes of urban catchments, coupled with limited availability of measured data, limits the prediction accuracy of conventional models. Artificial Neural Network models (ANNs) have displayed commendable progress in recognising and simulating highly complex, non-linear associations allied with input-output variables, with limited comprehension of the underlying physical processes. Therefore, this paper investigates the effectiveness and accuracy of ANN models, in estimating the urban catchment runoff, employing minimal and commonly available hydrological data variables – rainfall and upstream catchment flow data, employing two powerful supervised-learning-algorithms, Bayesian-Regularization (BR) and Levenberg-Marquardt (LM). Gardiners Creek catchment, encompassed in Melbourne, Australia, with more than thirty years of quality-checked rainfall and streamflow data was chosen as the study location. Two significant storm events that transpired within the last fifteen years - the 4th of February 2011 and the 6th of November 2018, were nominated for calibration and validation of the ANN model. The study results advocate that the use of the LM-ANN model stipulates accurate estimates of the historical storm events, with a stronger correlation and lower generalisation error, in contrast to the BR-ANN model, while the integration of upstream catchment flow alongside rainfall, vindicate for their collective impact upon the dynamics of the flow being spawned at the downstream catchment locations, significantly enhancing the model performance and providing a more cost-effective and near-realistic modelling approach that can be considered for application in studies of urban catchment responses, with limited data availability.