Expanding the use of ensemble streamflow forecasts in Australia

This thesis seeks to remove some of the technical barriers to the implementation and use of operational ensemble streamflow forecasting systems in Australia. There are two main parts to the thesis: the first addresses long-range (1-12 month) forecasts, and the second addresses short-medium range (0-10 day) forecasts.

There are currently no 12-month streamflow forecasting services in Australia. Water managers typically use ensembles sampled from historical data for planning purposes. These climatology ensembles are statistically reliable, but uninformative. In part 1 of the thesis we propose a new method, Forecast Guided Stochastic Scenarios (FoGSS), to generate ensemble streamflow forecasts at the monthly time step to 12 months, as an informative alternative to climatological ensembles.

Chapter 2 describes the theoretical framework for FoGSS: it is designed to combine calibrated climate forecasts with a hydrological model and error model. Uncertainties from climate forecasts and hydrological modelling are treated separately. Climate forecasts from coupled climate predictions systems are calibrated with an existing calibration method. This thesis develops a new error model within FoGSS for the treatment of hydrological uncertainties. The FoGSS error model combines data transformation, a monthly varying bias-correction, an autoregressive model and monthly varying Gaussian noise in a series of stages. We show the FoGSS error model effectively reduces forecast errors and reliably quantifies forecast uncertainty in two test catchments. When added to the calibrated climate forecasts, FoGSS produces reliable ensemble forecasts that are skillful at short lead times and transition to climatology-like forecasts at longer lead times.

Chapter 3 applies FoGSS to 63 catchments across Australia, including ephemeral rivers. We show that in general skill in forecasts derives from hydrological initial conditions rather than the calibrated climate forecasts. We refine the FoGSS error model by i) testing three candidate hydrological models and ii) applying a Bayesian prior to the inference of the bias-correction parameter. The prior encourages FoGSS to return climatology forecasts in the absence of strong evidence of forecast skill. We find that in general FoGSS performs well across a range of catchments, producing reliable ensemble forecasts that are skillful at short lead times and transit to climatology-like forecasts at long lead times. The GR2M hydrological model and the prior on the bias-correction improved forecasts. FoGSS did not perform well in highly ephemeral rivers, because the error model cannot produce >50% zero values.

Chapter 4 addresses the performance of the FoGSS error model in ephemeral catchments. We combine a technique to censor both simulation and observations with the FoGSS error model. The censoring method, and its implementation in forecast generation, allow FoGSS to generate >50% zeros and produce reliable forecasts in even highly ephemeral rivers. At the commencement of this thesis, no short-medium range ensemble streamflow forecasting system existed in Australia. Ensemble numerical weather prediction forecasts, as well as post-processing methods, are relatively well established, so Part 2 of this thesis concentrates on hydrological and error modelling.

Ensemble streamflow forecasting systems are often highly automated and require long observation records to calibrate hydrological models. Daily records of rainfall are reasonably common in Australia, however long records of hourly rainfall data are less so. This limits the regions in which short-medium term forecasting services can be quickly established. In Chapter 5, we demonstrate that the simple expedient of using disaggregated daily rainfall in calibration can produce robust hourly hydrological model parameter sets. In addition, we compare several hourly conceptual hydrological models, finding that GR4H performs better than the other models trialled.

In Chapter 6, we address a neglected topic: the propagation of hydrological forecast uncertainty to short-medium forecast horizons. Error models can dramatically increase the accuracy - and thus utility - of streamflow forecasts issued at gauges. However, conventional error models are often developed to produce predictions a single lead time in advance. We build on an existing error model - Error Reduction and Representation in Stages, ERRIS -¨ to improve the propagation of uncertainty to many lead times (>150 h time steps). We develop a moving average bias-correction and alter the restriction on autoregressive updates to improve the propagation of uncertainty through lead times and downstream. The restriction plays an important role in mitigating over-corrections by the autoregressive model, especially for floods. The revisions to ERRIS produce more reliable estimates of uncertainty at long lead times. The improvements are strongest in an ephemeral river, where ERRIS previously performed poorly.

Flood forecasts are a vital application of ensemble streamflow forecasts. However, ensemble flood forecasts are more complicated to verify than deterministic forecasts, which we address in Chapter 7. A key point of difference is the assessment of the reliability of ensemble flood forecasts, which is not possible for deterministic forecasts. To assess reliability of flood forecasts, only a small subset of forecasts from any forecast archive are relevant: those that are likely to coincide with floods. Previous studies have shown that selecting such a subset must be performed only on information contained in the forecasts. We develop two methods for performing this selection and demonstrate that these methods produce markedly different results to selection based on observed floods. In addition, we develop a new diagnostic we call 'Zappa plots' to measure the reliability of the magnitude and timing of ensemble forecasts of flood peaks.

The methods developed in this thesis have been encoded in operations-grade software for use in operational forecasting systems. Our philosophy of separating climate/weather uncertainty from hydrological uncertainty allows for highly modular modelling chains, ideal for operational systems. Research conducted for Part 2 of this thesis has contributed to a new ensemble 7-Day streamflow forecasting service, while the FoGSS methods are now being operationalized for hydropower generation.