Evaluation of the Bayesian joint probability approach to seasonal streamflow forecasting for inflows into Melbourne Water and Hydro Tasmania storages

[摘要] Skilful seasonal streamflow forecasts are potentially valuable to a wide range of users.The Bureau of Meteorology has adopted the Bayesian joint probability (BJP) modelling approach to seasonal streamflow forecasting to underpin a seasonal streamflow forecasting service.An experimental service has been operating since December 2009, producing forecasts for catchments in the southern Murray Darling Basin. This study has applied the BJP modelling approach to two additional geographical regions: Melbourne Water reservoir inflows and Hydro Tasmania hydroelectric scheme inflows.Forecasts of three-month inflow totals into the four main Melbourne Water reservoirs investigated in this study are predominantly of low to moderate skill (Root Mean Squared Error in Probability (RMSEP) skill scores between 10 and 40%), with RMSEP skill scores greater than 40% being achieved in one season for two sites.These forecast skills are substantially higher than the skill of seasonal rainfall forecasts and are potentially of some value to water managers. Most of the forecast skill is derived from indicators of the initial catchment conditions.Predictors representing future climate influences are only selected for forecasts made between August and November.Indices representing surface temperature anomalies in the Indian Ocean dominate the selected predictors representing future climate influences. Forecasts of three-month inflow totals into the two Hydro Tasmania hydroelectric schemes investigated in this study have little skill and are unlikely to be useful for practical application.RMSEP skill scores exceed 10% for only two months in the Mersey-Forth hydroelectric scheme and not at all for the West Coast catchments.The lack of skill in the forecasts of inflows into the Tasmanian hydroelectric schemes arises because there is little persistence in streamflows and rainfall is very difficult to forecast using statistical methods.Overall the majority of forecasts at all sites appear to be reliable and have little bias.However, forecasts for seasons that include autumn do appear to have a temporal bias, with few observations falling into the upper tail of forecast distributions after 1980.This means that the autumn streamflow forecasts after 1980 tend to be higher than the observed values. This bias is consistent with the observed decline in autumn rainfall with has been attributed to the strengthening of the subtropical ridge.Future research will investigate the importance of predictors that describe the subtropical ridge to seasonal streamflow forecasting.The skill of seasonal streamflow forecasts is highly dependent on the hydrological memory of catchments.Where the catchment hydrological memory is high, including predictors representing initial catchment conditions into forecasting models will produce forecasts with some useful skill.Where the catchment hydrological memory is low, for example in Tasmania, the production of skilful streamflow forecasts is dependent on having skilful forecasts of rainfall.Improving the skill of seasonal streamflow forecasts will require skilful forecasts of rainfall. In this study, rainfall forecasts were produced jointly with streamflow forecasts forecasts using the BJP modelling approach. The skill of these rainfall forecasts is low.Seasonal rainfall forecasts using the Bureau of Meteorology’s POAMA dynamical climate model appear to have some skill for the catchments considered in this study.Future research will indentify how best to use the output from POAMA to increase the skill of seasonal streamflow forecasts throughout Australia.

[发布日期] 2010-12-31 [发布机构] CSIRO

[学科分类] 地球科学(综合)