Improving error modelling and calibration of continuous hydrological models for streamflow and flood forecasting

[摘要] This report documents our attempt to improve the System for Continuous Hydrological Ensemble Forecasting (SCHEF) by:1.Investigating hydrological model calibration objectives with the aim of improving flood forecasts2.Developing improved error models that reliably quantify the error in streamflow and flood forecastsOf the hydrological model calibration objectives tested, we find that the Kling-Gupta Efficiency slightly outperforms other calibration objectives for flood forecasting applications. However, we conclude that improvements in flood forecasts gained from calibration objectives are likely to be small compared to improvements gained from better rainfall forecasts.The error model currently employed in SCHEF showed a major deficiency for flood forecasting: the occasional tendency to grossly over-correct receding streamflows. We develop a restricted autoregressive error model applied to normalised data (RAR-norm) that mitigates this and other deficiencies inherent in autoregressive error models. The RAR-norm model performs well at simulating flows at a daily time step. More rigorous testing of the RAR-norm model at the hourly time step, and in flood forecasting applications, is planned for the 2014/2015 financial year.One of the major benefits of ensemble streamflow forecasts is the formal quantification of forecast uncertainty. Previous work showed that the forecast uncertainty estimated by SCHEF could be improved, in particular at the shorter lead times most pertinent to flood forecasters (2 days). To address this issue we couple the RAR-norm approach with a Gaussian mixture distribution to model hydrological errors. We estimate parameters for these models with a method we call the multi-stage approach. In the multi-stage approach, the parameters of the hydrological models and different components of the error model are estimated in stages. This minimises interference between parameters, ensures each model performs its function as it is designed to, and minimises computation. The multi-stage approach results in an error model that is highly effective at describing errors at the daily time step. In this study the multi-stage approach is tested largely on daily streamflows, with some preliminary analyses of hourly streamflows. These analyses show that using the multi-stage approach to estimate parameters in the RAR-norm model with the Gaussian mixture distribution has the potential to be a useful replacement to the error model currently implemented in SCHEF. The multi-stage approach will be further developed and verified in the 2014/2015 financial year.On the basis of this work, we conclude and recommend the following:1.The performance of flood forecasts did not change markedly with different calibration objectives. However, KGE outperformed other objectives sufficiently frequently, and has sufficient theoretical advantages, that we recommend its use as an objective for calibrating hydrological models for continuous flood forecasting.2.Calibrating hydrological models to observations above thresholds (the median and 75th percentile flows) did not improve the simulation of floods, and we do not recommend this approach.3.There is probably greater scope for improving flood forecasts by using better rainfall forecast information than through further refinement of hydrological model calibration objectives, and we recommend that future research be focussed accordingly.4.The error model currently used in SCHEF can seriously distort forecast flood hydrographs. A better error model is required.5.There is a good theoretical basis for applying error models to transformed (normalised) streamflows, rather than correcting untransformed flows (as currently implemented in SCHEF).6.The performance of the RAR-norm error model at the daily time step is sufficiently promising that it should be considered a leading candidate as the basis of a new error model for SCHEF. Additional testing of the RAR-norm model at the hourly t

[发布日期] 2014-09-04 [发布机构] CSIRO

[学科分类] 地球科学(综合)