Better accounting of droughts in long-range inflows scenarios with TULIP

Many water management agencies rely on stochastic inflow scenarios to plan water operations. For example, Hydro Tasmania, Australia’s largest hydropower generator and water manager, relies on 20+ year inflow scenarios to assess the long-range sustainability of their power generation system. A variety of methods are available for stochastic data generation, but many assume a stationary climate. In locations where inflow has long-term trends, assuming a stationary climate in stochastic data generation is likely to underestimate future wet or dry extremes, in particular for sequences of dry or wet months or years.

To address this issue, we have developed the Trend and Uncertainty in Long Inflow Predictions (TULIP) model. TULIP is a Bayesian model that generates long-range predictions of inflows at the monthly time step. TULIP accounts for:

• Heteroscedasticity and skew in inflow data by using data transformation with the sinh-arcsinh transformation, and zero values with censoring
• Spatial correlation between inflow sites
• Autocorrelation using a first-order autoregressive model
• Linear trend in inflow
• Seasonal variation in properties (1)-(4), using Fourier series to control the parameters

TULIP is being implemented operationally by Hydro Tasmania to replace its existing method of generating stochastic scenarios, which assumes a stationary climate. At sites with long-term trends in historical inflow, we show that TULIP produces more reliable long-range predictions than is possible if a stationary climate is assumed. This allows TULIP to produce sharper ensembles and more realistic projections of future drought, allowing Hydro Tasmania to better plan for the long-range sustainability of its system. In this presentation we describe the TULIP model and its performance. We also discuss future plans to incorporate information on inflow trends from global and regional climate models into TULIP.