Comparison between Perfect Foresight and Myopia/Timestep-Solution

[Alexander Burkhardt]

Hi Community,

I am looking to find people who have dealt with this sort of issue/modelling who would be interested to exchange some ideas and talk about the methodology.
My goal is to run the same Model under different lengths of Foresight and compare those results to the results under Perfect Foresight.

So far, I have been successful in implementing some Timestep-scenarios (20, 15 and 10 year windows. However, 5 Year windows cause infeasibilities) in our big model. The differences between the Model solutions are, however, rather small. For example, total costs differ only by 0,001 % or so. FEC is almost the same across all sectors. I expected a lot more differences.

Then I implemented timestep-scenarios in the TIMES Demo Model, just as an exercise. Here, even 5 years worked. Again, there are at best minimal differences. The only real difference which I saw was caused by a User Constraint (for the 5 year timestep), thus leading to dummy imports. Removing the UC led to the same results basically, regardless of the level of foresight.

Therefore, it would be great to talk to people who have experimented with Myopia and would be open to some informal exchange/chatting.

There are two questions in my head:
·        Can the usage of timesteps be offset by some Model-Feature, in a way that makes its impact much (much) smaller?
·        Are there known pitfalls when solving TIMES in a timestep-manner that would explain that kind of (non-)results?
 
Thanks!

[Antti-L]

I think what you have seen in your tests looks great, if some firm conclusion could be drawn from them:

1) Your tests would seem to suggest that, in general, energy-system related policies and technology changes would need to be well known only about 10 years ahead without major losses to businesses, because the results show only a minor economic loss when any further-ahead information is lost.

2) Your tests would seem to suggest that a recursive-dynamic solution can in (most?) cases reach a solution that is not far apart from the perfect foresight solution, if the algorithm is carefully implemented. And therefore, one might save quite a lot in model solution time by using a recursive-dynamic solution (as e.g. is more or less standard in the Balmorel model).

However, you did not describe much about the test cases, and I suspect they may not well represent cases where strong system transitions are needed.

>  Are there known pitfalls when solving TIMES in a timestep-manner that would explain that kind of (non-)results?

Are you calling results that are close to the perfect foresight solution (even when assuming limited foresight) "non-results"? Do you mean that some kind of pitfalls might somehow reduce the losses in the perfect foresight information when using the time-stepped solution?

[olexandr]

Very interesting discussion! I would be happy to contribute / participate.

I'll try to find and share some comparisons using Balmorel (the model Antti referred to). There has been a number of them made over the years, but I don't recall whether any of them were published.

This report maybe useful: https://www.researchgate.net/publication...tice_Guide

Regarding the "non-results", in my view, theya are indicative of how "locked" the model is. However, I agree with Antti that it is impossible to say much without knowing more about the test cases.

[olexandr]

There is some discussion and comparison using Balmorel (e.g. from page 178) here: https://backend.orbit.dtu.dk/ws/portalfi...Thesis.pdf

[Antti-L]

Thanks for the interesting links, Olexander.

Some additional older links to anyone who might be interested and have missed these:

  • IER implementation of Myopic TIMES: times_myopic.pdf
  • IER experience with the Pan-EU model: Blesl_Myopic.pdf
  • PSI case study with TIMES: Myopic_vs_clairvoyant.pdf
  • UKTM case study: Nerini_presentation.pdf
  • Balmorel case study: BalmorelInvestmentHorizonUncertainty.pdf

To my knowledge, the IER implementation of running TIMES with Myopia has not been made available to ETSAP, and therefore the insights from that work have unfortunately not been available for the current TIMES implementation.

P.S. Concerning the TIMES Demo model, where some UC contraint caused dummy imports when using 5 year steps, I'd be interested to see that model case, if available (I don't know which demo model variant this is). 

[Alexander Burkhardt]

Thanks for the interesting links, Olexander.

Some additional older links to anyone who might be interested and have missed these:

  • IER implementation of Myopic TIMES: times_myopic.pdf
  • IER experience with the Pan-EU model: Blesl_Myopic.pdf
  • PSI case study with TIMES: Myopic_vs_clairvoyant.pdf
  • UKTM case study: Nerini_presentation.pdf
  • Balmorel case study: BalmorelInvestmentHorizonUncertainty.pdf

To my knowledge, the IER implementation of running TIMES with Myopia has not been made available to ETSAP, and therefore the insights from that work have unfortunately not been available for the current TIMES implementation.

P.S. Concerning the TIMES Demo model, where some UC contraint caused dummy imports when using 5 year steps, I'd be interested to see that model case, if available (I don't know which demo model variant this is). 

Hello Oleksander and Antti,

thank you very much for your valuable input. I very much appreciate it.

Maybe I did not give enough background. I am doing my PhD at IER on this topic. The goal was, similiar to the Nerini Paper referenced here, to use myopic modelling in the context of CO2-pricing shemes and policy interventions overall, as those seem to be most contigent on a sufficient level of foresight.

I don't think that 10 years would be necessary enough for policy changes, at least in some cases. However, in reality, we see actual foresight being more like 5 years or so, sometimes even less (at least in Germany). And I would expect higher costs compared to perfect foresight nontheless.
However, one important point that I have gathered from Oleksanders Link is that the discount rate matters a lot. I thought about modifying it previously, but have not done it so far.

Another thing is, I used no CO2-Bounds, but CO2-Prices/Taxes. These had a signficant hike (x4 in one Period) after 2030. In this case, I should be able to observe some changes in the 2015 and 2020 periods for technologies with lifetimes > 15 years. But so far, I was not able to notice anything significant. Actually, if I was able to spot a difference, it would go against my expectations (e.g. lower emissions before CO2-pricing starts), which does not make sense to me.

Regarding the Demo Model, I have attached the file names. It was in my Veda Folder, but outside the Demo Models used for the documentation and training. I choose this one as it is fairly simple, but has a carbon tax. Antti, if you need more information, I could also send it to you.

I will post some updates here on further developments.