The data-based mechanistic (DBM) modelling methodology has previously been proposed for developing real-time flood forecasting models. However, only limited research has been published comparing the performance of DBM models with commonly used conceptual rainfall runoff (CRR) models and investigating the predictive performance of DBM models when transferred in time to conditions outside the range of their calibration data.

This thesis reports a case study performance comparison of DBM and CRR rainfall runoff models, including temporal transfer, across a range of forecasting performance measures. The DBM models were based upon the range of model structures and parameter estimation approaches used in other published studies, and did not attempt to exhaustively explore the potential of the DBM methodology. The principal findings are that the employed DBM models could not improve upon the CRR models and that neither model type performed well when transferred in time to conditions outside the range of their calibration data.

The performance comparison was followed-up with a synthetic catchment study to investigate the role of model (as opposed to data) errors on the predictive performance of DBM models when transferred in time to conditions outside the range of their calibration data. This work found that the DBM modelling approach as applied here cannot be relied upon to estimate models that will transfer satisfactorily in time, even with error-free data, indicating that shortcomings in temporal transferability are intrinsic to the modelling process. The work further found that calibration performance was not a reliable predictor of temporal transferability amongst a set of identified candidate models.

The flexibility of DBM methodology does, however, make it suitable for incorporating future developments, aimed at mitigating the identified shortcomings. Potential areas of investigation include the use of alternative model structures and optimisation schemes, as well as the use of hybrid DBM-CRR models.