Allow dependency loops between commodities

[alexdewar]

Users are currently prevented from creating dependency loops between commodities (see #519), where one commodity is necessary for the production of another -- directly or indirectly -- which produces the original commodity. For example, you may have a hydrogen power plant which produces electricity but also requires electricity as an input in order to operate. @ahawkes has reported that such dependency loops are quite common in the real world. Apparently MUSE1 supports this functionality, though according to @tsmbland the results are often surprising, so we may not want to imitate the MUSE1 approach (insofar as this would be possible anyway, given that the model is different).

We would presumably want to adopt some kind of iterative approach where, to go back to the previous example, we would first run dispatch to calculate the electricity demand of the hydrogen plant, then rerun dispatch to see the effect of the new generation of electricity on the system. We should figure out what to do in the case that the system doesn't converge on a stable state (probably just have an upper limit on number of iterations).

On the technical side, we use the petgraph crate for representing and analysing the dependencies between commodities (in addition to checking for loops, it is also used for checking other invariants to do with commodities hold). We do a topological sort to figure out the order in which investment should be performed, which also checks for cycles in the process. We'd have to figure out how to adapt this code (petgraph seems pretty comprehensive, so I'm sure there's a way).

[tsmbland]

Apparently MUSE1 supports this functionality, though according to @tsmbland the results are often surprising, so we may not want to imitate the MUSE1 approach (insofar as this would be possible anyway, given that the model is different).

To clarify, MUSE1 "supports" this in the sense that it will often be able to run to completion even if there are loops, but this violates a lot of the assumptions in the model, and there's no explicit support for loops. So in this sense there's no "MUSE1 approach" for us to imitate

[ahawkes]

circularity.zip

Here is a model that includes a circularity.

Based on this diagram:

Example schematic - circularity.pptx

[tsmbland]

@ahawkes I've tried running your model (with a few small adjustments - #914), and I can get it to pass the dispatch run in the first year. See attached results files. There are a few differences in the results compared to the two_outputs model, which it's identical to apart from the hydrogen loop. Notably some lower prices, which might be a good sign.

I think the main unresolved challenge with loops is agent investment, where our current approach very much assumes a directed acyclic graph of commodity flows. I'm going to open a separate discussion about this because I have some ideas.

circularity.zip

Edit: results with debug files circularity.zip