Towards dynamic catchment modelling: a Bayesian hierarchical mixtures of experts framework |
| |
| Authors: | Lucy Marshall David Nott Ashish Sharma |
| |
| Institution: | 1. School of Civil and Environmental Engineering, The University of New South Wales, Sydney, Australia;2. School of Mathematics, The University of New South Wales, Sydney, Australia |
| |
| Abstract: | Despite the abundance of existing hydrological models, there is no single model that has been identified as performing consistently over the range of possible catchment types and catchment conditions. An attractive alternative to selecting a single model is to combine the results from several different hydrological models, thereby providing a more appropriate representation of model uncertainty than is the case otherwise. Methods based on Bayesian statistical techniques provide an ideal means to compare and combine competing models, as they explicitly account for model uncertainty. Bayesian model averaging is one such alternative that combines individual models by weighting models proportional to their respective posterior probability of selection. However, the necessity of having fixed weights for each model over the entire length of the simulation period means that the relative usefulness of different models at different times is not considered. The hierarchical mixtures of experts (HME) framework is an appealing extension of the model averaging framework that allows the individual model weights to be estimated dynamically. Consequently, a model more capable at simulating low flow characteristics attains a higher weight (probability) when such conditions are likely, switching over to a lower weight when catchment storage increases. In this way, different models apply in different hydrological states, with the probability of selecting each model being allowed to depend on relevant antecedent condition characteristics. HME models provide additional flexibility compared with simple combinations of models, by allowing the way that model predictions are combined to depend on predictor variables. Thus, for hydrological models, the ‘switch’ from one model to another can depend on the existing catchment condition. This new modelling framework is applied using a simple conceptual model to 10 selected Australian catchments. The study regions are chosen to vary considerably in terms of size, yield and location. Results from this application are compared with the alternative where a single fixed model structure is applied. Comparison of the model simulations using the maximum log‐likelihood and the Nash‐Sutcliffe coefficient of efficiency show that more variance in streamflow was explained by the HME model, compared with the conceptual model alone for each of the catchments investigated. Copyright © 2006 John Wiley & Sons, Ltd. |
| |
| Keywords: | Bayesian rainfall‐runoff model model averaging |
|
|
