Group for Research in Decision Analysis


Modeling Uncertainty in a Large Scale Integrated Energy-Climate Model

, , and

The well-known method stochastic programming in extensive form is used on the large scale, partial equilibrium, technology rich global 15-region TIMES Integrated Assessment Model (ETSAP-TIAM), to assess climate policies in a very uncertain world. The main uncertainties considered are those of the Climate Sensitivity parameter, and of the rate of economic development. In this research, we argue that the stochastic programming approach is well adapted to the treatment of major uncertainties, in spite of the limitation inherent to this technique due to increased model size when many outcomes are modeled. The main advantage of the approach is to obtain a single hedging strategy while uncertainty prevails, contrary to classical scenario analysis. Furthermore, the hedging strategy has the very desirable property of attenuating the (in)famous ‘razor edge’ effect of Linear Programming, and thus to propose a more robust mix of technologies to attain the desired climate target. Although the example treated uses the classical expected cost criterion, the paper also presents, and argues in favor of, altering this criterion to introduce risk considerations, by means of a linearized semi-variance term, or by using the Savage criterion. Risk considerations are arguably even more important in situations where the random events are of a ‘one-shot’ nature and involve large costs or payoffs, as is the case in the modeling of global climate strategies. The article presents methodological details of the modeling approach, and uses realistic instances of the ETSAP-TIAM model to illustrate the technique and to analyze the resulting hedging strategies. The instances modeled and analyzed assume several alternative global temperature targets ranging from less than 2oC to 3oC. The 2.5oC target is analyzed in some more details.

The paper makes a distinction between random events that induce anticipatory actions, and those that do not. The first type of event deserves full treatment via stochastic programming, while the second may be treated via ordinary sensitivity analysis. The distinction between the two types of event is not always straightforward, and often requires experimentation via trial-and-error. Some examples of such sensitivity analyses are provided as part of the TIAM application.

, 26 pages