Revise equations for learning rate improvements in endogenous learning formulas

[robbieorvis]

As part of my review of the revenue assignment for BEVs, I started looking at our battery price projections in the EPS in the endogenous learning sector. Based on a review of the input data (including source files) and our results, our prices fall far too fast, at least for batteries and possibly for other technologies, especially with updated data on global production estimates for batteries.

We currently use a learning rate of 23.5%, taken from this paper: https://pubs.rsc.org/en/content/articlelanding/2021/ee/d0ee02681f#eqn2

Here is where other studies are landing on projected battery costs. See this chart from ICCT for confirmation:

When I used the updated data on projected battery production values with the current approach in the model, I get battery prices that are far too low, falling to <$30/kWh by 2030 and all the way to <$13/kWh in 2050.

We should consider revisit the calculations and data here to understand what is going on.

[robbieorvis]

Okay, @dobrien13 found a formula error that negates all the math I did in terms of the fix and the error, but there is still something going on with the fact that our prices are so far off from what many other groups are publishing. I think we should look into why that's happening some more.

[jrissman]

For technology prices that support endogenous learning (like wind, solar, batteries, etc.), we have a flexible structure right now, which lets you customize:

• The endogenous learning rate, i.e., how much prices drop per unit of technology deployed in the modeled region
• The exogenous price decline schedule
• The share to which the learning rate is endogenous vs. exogenous, allowing for any blend of these two methods (100% endogenous, 100% exogenous, or anything in between)

We came up with this customizable approach because purely endogenous learning was overstating the effect of deployment in a single country on global technology prices, and was also missing how deployment of a technology outside the modeled region borders can lower the price of that technology within the modeled region. On the other hand, fully exogenous pricing didn't allow us to simulate the way that policies that boost technology deployment can help drive technologies down their learning curves. Our view was that a blend of the two approaches produced the best results.

It's all data-driven, so you should be able to mirror any learning curve you find in an external source by selecting your endogenous learning rate, your own exogenous price schedule, and your own balance between them. (In fact, there are infinite ways to mirror the external source, which range from being more or less sensitive to policies within the region, so you would need to calibrate both the learning curve and how dramatically you want policies to be able to bend the curve. Policies in China should bend the curve a lot, while policies in Rhode Island should bend it very little.)

Should this issue be tagged as "data-only"? Or have you encountered some sort of problem that cannot be solved by customizing the input data as described above?

I mean, other than the issue of non-battery EV components, since that one has its own GitHub issue (#267).

[dobrien13]

note: come back to this next week prices match other projections if using a 2022 base year and "all-battery-type" learning rate from Ziegler & Trancik

[dobrien13]

Learning rate should be 18.9% for cumulative market size measured in energy capacity.

[robbieorvis]

This is very tricky, in part because COVID caused a disruption in the supply chain that has caused battery prices to go up in the last year or two, though that trend should mitigate.

I think the most likely thing may be some issue related to historical production of batteries that is missing from our file and so our rate of decline is larger than it should be. I am still working on this and will report back. The upshot is that using the existing data + learning rates (even modified to reflect our different read of the data source) + latest data on projected battery deployment results in battery price estimates that are too low, even when it is moved to 100% global (meaning the in region deployment does nothing. Basically just taking our learning rate approach and published learning rates and projected future battery manufacturing, we are too high. I will keep digging and circle back.

[robbieorvis]

I found the issue: there was a formula error that didn't include any of the estimated battery production capacity historically, so we were essentially starting from much earlier on the learning curve. Once fixed, that resolved this issue. It was indeed a data only fix and we will incorporate the data fix into the next model update.