Towards a recursive bayesian total error analysis framework

Abstract

The Bayesian total error analysis (BATEA) framework seeks to provide an improved description of the uncertainties affecting environmental modelling through the use of user defined explicit error models describing input, output and model uncertainty. This allows a more informed assessment of model performance and predictive ability. BATEA has seen application to a range of rainfall-runoff and river basin models. A significant limitation of the current applications in their reliance on batch processing of data. In batch calibration, when input and model errors are treated as latent variables, the dimension of the parameter space grows with record length. This limits batch calibration to relatively short record lengths and makes real-time applications involving forecasting impractical. This study targets this problem by developing a recursive implementation of BATEA based on particle filters. Recursive estimation can be considerably faster because the parameter space at each time step is small compared with the batch space. This study shows how the particle filtering technique Sequential Importance Sampling (SIS), traditionally used in automatic control and signal processing applications, can be adapted to the BATEA framework.