Reliable hourly streamflow forecasting with emphasis on ephemeral rivers

Abstract

Abstract Ephemeral rivers pose considerable challenges for hydrological forecasters. A staged error model is devised for hourly streamflow forecasting with emphasis on ephemeral rivers. Apart from key components such as bias-correction, autoregressive updating and mixed Gaussian residual distributions, the model treats both observed and simulated zero flows as censored data. This assumption requires additional novel parameter estimation and ensemble generation methods. When inferring error model parameters by maximum likelihood estimation (MLE), the likelihood is derived to account for simulated and observed streamflow taking zero or non-zero values using data censoring. MLE is trialed to be replaced by least-squares-after-transformation (LST) estimation wherever possible to ease computational burdens. When generating ensemble forecasts, a data augmentation is used approach to recover censored data. The staged error model is applied to 18 ephemeral and 11 perennial rivers in Australia and generate hourly streamflow forecasts up to 10 days. LST leads to similar or even marginally better forecast performance than MLE by checking with various verification metrics, including bias, continuous ranked probability score, the α -index and the average width of confidence intervals. Skillful and reliable ensemble forecasts can be generated in 25 out of 29 catchments even at extended lead times. However, propagation of uncertainty can sometimes be challenging in four highly ephemeral catchments, leading to unreliable forecasts at long (e.g. >40 hour) lead times. Prospects for addressing this issue in future is discussed.