This report assesses the value of 'end-to-end' forecasting of economic variables that have a weather-dependent component, using the ensemble prediction system (EPS) of the European Centre for Medium-range Weather Forecasting (ECMWF) [1]. It considers decision-making scenarios in which the outcome is a continuous variable (e.g. demand) and in which the decision (e.g. production) is also a continous variable. Making binary choices (e.g. to salt or not salt potentially icy roads) is well described by cost-loss analysis [2]. The current report aims to move away from the cost-loss ratio toward questions of 'how much?' rather than questions of 'whether or not?'. The basic concepts of utility maximization in decision making are outlined with an idealized example of bagel sales. These concepts are then applied to the problem of forecasting electricity demand in 12 cities around the world. They are also applied to forecasting potential wind generation at a site in the UK. The period from January 1999 to October 2000 was used to test the value of the ECMWF ensemble prediction system to decision makers in these two scenarios. It was found that the ECMWF EPS yields more optimal decisions in both scenarios. The results indicate that it is essential to treat all deterministic forecasts (control or ensemble members) as probabilistic forecasts, by attempting to estimate the error distribution about each forecast. It was also found that the 51 member ensembles are not distinguishable from a 'singleton ensemble' distribution constructed by adding historical errors to the best guess (high resolution control) forecast on approximately two-thirds of the days in the period. However, on the days on which the ensembles are distinguishable on average they outperform the best guess forecast combined with the historical errors. A crude method for estimating the size of the ensemble required to increase the number of days on which the ensemble would be distinguishable was applied to the forecasts in the test period. It was found that, for over 95% of ensemble forecasts to be distinguishable, ensembles of at least about 100 members would be required.
End to end ensemble forecasting: Towards evaluating the economic value of the Ensemble Prediction System
von Hardenberg;
2001
Abstract
This report assesses the value of 'end-to-end' forecasting of economic variables that have a weather-dependent component, using the ensemble prediction system (EPS) of the European Centre for Medium-range Weather Forecasting (ECMWF) [1]. It considers decision-making scenarios in which the outcome is a continuous variable (e.g. demand) and in which the decision (e.g. production) is also a continous variable. Making binary choices (e.g. to salt or not salt potentially icy roads) is well described by cost-loss analysis [2]. The current report aims to move away from the cost-loss ratio toward questions of 'how much?' rather than questions of 'whether or not?'. The basic concepts of utility maximization in decision making are outlined with an idealized example of bagel sales. These concepts are then applied to the problem of forecasting electricity demand in 12 cities around the world. They are also applied to forecasting potential wind generation at a site in the UK. The period from January 1999 to October 2000 was used to test the value of the ECMWF ensemble prediction system to decision makers in these two scenarios. It was found that the ECMWF EPS yields more optimal decisions in both scenarios. The results indicate that it is essential to treat all deterministic forecasts (control or ensemble members) as probabilistic forecasts, by attempting to estimate the error distribution about each forecast. It was also found that the 51 member ensembles are not distinguishable from a 'singleton ensemble' distribution constructed by adding historical errors to the best guess (high resolution control) forecast on approximately two-thirds of the days in the period. However, on the days on which the ensembles are distinguishable on average they outperform the best guess forecast combined with the historical errors. A crude method for estimating the size of the ensemble required to increase the number of days on which the ensemble would be distinguishable was applied to the forecasts in the test period. It was found that, for over 95% of ensemble forecasts to be distinguishable, ensembles of at least about 100 members would be required.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
