replace stationary by ONONSPEC projection

[robinhasse]

This PR introduces a new projection of non-specified IEA balance flows that are currently mapped to buildings. A different allocation across sectors proposed in #609 is not subject of this PR.

status quo

There is a very old file that has SSP1 - SSP5 scenarios for all demand mapped to industry and buildings. All energy demand that is mapped to buildings but not represented in EDGE-B currently comes from this file as "stationary". In convertStationary, the outdated historic demand from the file is replaced by current IEA balance data that linearly transitions to the projections in the file when going into the future. The respective flows belong to the IEA flow ONONSPEC - other non-specified demand. Even though we have no idea what this is, there are different projections either approaching zero or higher values. In the most striking example, we see exponentially increasing historic electricity demand in CHN but a decay towards zero in the projections.

change

This PR gets rid of the file overall. Instead we now project the IEA demand dynamically based on the most recent historic observations. As we have no idea, where the demand comes from, I want to avoid to make too specific assumptions for the future, so we just continue with the recent slope and then approach a constant value.

method

We assume an asymptotic model. It start from the level $x_0$ at the end of history (EOH) and with the EOH slope $\dot{x}_0$ and approaches the fraction $\epsilon$ of this slope within $\Delta t$. Both $x_0$ and $\dot{x}_0$ are determined through a linear regression of the last $n$ time steps.

$x(t) = x_0 + \dfrac{\dot{x}_0}{c} \cdot [ 1 - \exp ( - c \cdot ( t - t_0 ) ) ]$
with the decay rate $c = -\dfrac{\ln \epsilon}{\Delta t}$

The method requires the following parameters. I indicated my current choice

• $n = 5$: not as relevant, I think.
• $\epsilon = 0.05$: tolerance to considere the slope constant. The higher this parameter, the higher (lower) are approached long-term values for positive (negative) EOH slope.
• $\Delta t = 15 - 25 yr$: time span within which slope has decreased to a value considered constant: $\dot{x}(t_0 + \Delta t) = \epsilon \cdot \dot{x}_0$

scenarios

I had a hard time thinking of a scenario differentiation for the projections as we had them with stationary before. @robertpietzcker, do you think this is necessary? As a first example, I created a low, med and high scenario referring to the long-term projection level by differentiating $\Delta t$. We can adjust the parameterisation and would have to map the level scenarios to SSPs.

scenario	$\Delta t$, if $\dot{x}_0 > 0$	$\Delta t$, if $\dot{x}_0 \le 0$
low	15yr	25yr
med	20yr	20yr
high	25yr	15yr

This is a comparison of the projections showing the new scenarios as thick lines and the old ones in thin lines. Mind that the old scenarios are without my hot fix in #635.

implementation

This one is for @fbenke-pik: I currently get the relevant IEA flows through calcOutput("IOEdgeBuildings", subtype = "output_EDGE"). This gives me the stationary IEA flows in buildings. From these, I still filter only those flows that are not represented in EDGE-B. This works fine but seems a bit intransparent. It might be better to update the mapping structuremappingIO_outputs.csv to get just the flows we need here and not the ones represented in EDGE-B. Not sure if this would come with an overall change to mrcommons:calcIOEdgeBuildings and what cache implications would be.

[ricardarosemann]

Thanks, the implementation looks good and makes sense and I find most of the resulting projections reasonable (with few exceptions). Let me know if you want me to look at some more details.

[ricardarosemann]

Suggested change

	valueEOH = predict(m, list(period = periodEOH)))
	valueEOH = predict(m, data.frame(period = periodEOH)))

Not sure if this makes much of a difference, but usually predict takes data frames.

[robinhasse]

This is for $n=10$. Not radically different but outliers like in REF feothliquid distort less.

[robertpietzcker]

yes, the second version looks better, especially when looking at REF :-)
I like it

[robinhasse]

Thanks @robertpietzcker. If you could quickly answer these questions (or approve my suggestions). Thanks!

• Do you think, it is worth keeping a scenario differentiation in the projections?
• Do the 3 scenarios differentiated by long-term level make sense to you and are you happy with the parameterisation?
• How would you map the level scenarios to the SSPs? This would roughly fit total FE demand in EDGE-B scenarios:

SSP scenario	level scenario
SSP1	low
SSP2	med
SSP3	med
SSP4	med
SSP5	high

[robertpietzcker]

yes, looks good to me.
you could also map "high" to SSP3, as Nico is designing it to lead to high demands if I understood correctly
as you like

[robertpietzcker]

I can't comment on the R code, but the results look good