A model is written at one resolution and solved at another more often than not. The resolution may be reduced to fit a solver, to explore a design before committing to an overnight run, or to answer a question that the finest resolution obscures. energyRt offers several routes to do this, and they are not interchangeable: each changes a different thing, and the results of each mean something different.
This article maps the routes. Two companion articles cover one axis
each in depth, vignette("time-resolution") and
vignette("space-resolution"), with
vignette("region-aggregation") working through a full
spatial example.
Resolution is not the only way to make a model smaller, and the
routes below are not the only ones that do.
vignette("model-size") maps the same ground by cost rather
than by axis: it adds the horizon reductions and the drivers that bound
decisions instead of removing cells, and it sorts every route by whether
the answer is still the full model’s.
Four operations, often confused
Four distinct things are frequently described with the same words.
Subsetting keeps part of a model and discards the rest. The result is a smaller model of a smaller system, and it answers questions about that part.
Coarsening keeps the whole system but describes it with fewer regions or fewer timeslices. The result is a smaller model of the same system, and its answer approximates the finer model’s in a direction that is not always predictable.
Sampling keeps the whole system and the whole year, but represents the year with a subset of its periods, reweighted so that the sample still stands for a full year. Each sample estimates the same quantity; samples are replicates.
Decomposition leaves the model alone and splits the solve, recombining afterwards. Whether the parts add up depends on what was split.
The last distinction causes the most trouble in practice. Regional
sub-problems solved without trade have objectives that
sum to the full problem’s. Timeslice samples have
objectives that must not be summed, because each
already represents a whole year; they form a distribution, and
sample_summary() is the way to read it.
Space
| Operation | Verb |
|---|---|
| Keep a sub-territory | subset_model_regions() |
| Keep a sub-territory, during interpolation |
interpolate_model(geoscale = ) with a filtered
geoscale |
| Coarsen to a declared level | aggregate_model_regions(level = ) |
| Coarsen into derived groups | aggregate_model_regions(clusters = ) |
| Inspect candidate groupings |
get_process_groups(),
process_cluster_sweep()
|
| Split the solve by region | solve_by_region() |
Subsetting cuts a model down to the regions named. Trade routes with
one endpoint outside the kept set are replaced by boundary stubs, an
import or an export carrying the price at which the severed neighbour
would have traded. Without a price the route is dropped instead, which
makes the sub-model an island and generally understates its options.
boundary_window() derives a plausible price band from the
model’s own demand.
library(energyRt)
topia_elc <- topia$modules$electricity
mod <- newModel("TOPIA", data = topia_elc$repo,
calendar = calendars[["d365_h24_1dpm"]],
region = topia_elc$regions,
horizon = topia$modules$horizons$BY2025_50_by10,
discount = 0.05)
mod <- setGeoscale(mod, topia_elc$geoscale)
west <- subset_model_regions(mod, region = c("R01", "R02", "R03"))Aggregation is a different operation: no region leaves the system, but several become one. Every parameter of every object must then be combined, and how it is combined depends on what the parameter means. Extensive quantities — capacities, stocks, availability volumes, demands — are summed. Intensive ones — efficiencies, costs per unit, availability factors — take a mean weighted by the size of each contributing region.
Aggregation bounds the finer answer in neither direction. With uniform efficiency across the merged regions the coarse objective is lower, because merging removes transport limits. With heterogeneous efficiency it is higher, because averaging destroys the ability to dispatch the best unit first. Both directions are demonstrated in the test suite.
zones <- aggregate_model_regions(mod, level = "zone")A clustering is the same operation over groups derived from data
rather than declared, and the clustering itself is a scale-construction
method supplied by nestedscales. Because the groups are
derived, the sweep showing how quality varies with the number of groups
is reported rather than resolved: energyRt does not choose a number of
clusters.
Time
| Operation | Verb |
|---|---|
| Move a model onto another calendar | recast_model_calendar() |
| Draw timeslice samples | calendar_samples() |
| Solve each sample | solve_by_sample() |
| Split the horizon |
horizon_windows(), solve_recursive()
|
| Guide an expensive solve with cheap ones | solve_guided() |
Recasting moves every timeslice-indexed value onto a new calendar. Coarsening takes a weighted mean, which conserves the annual total exactly; refining copies the coarse value onto each finer slice, which conserves nothing but invents nothing either. The rule follows from the equations: a timeslice-indexed value is a rate, multiplied by the slice’s weight wherever it enters a cost or a cumulative total.
annual <- recast_model_calendar(mod, topia$modules$calendars$annual)A recast model is not a smaller answer to the same question. Its dispatch, its costs and its objective are its own, and comparing them against the fine model’s compares two models rather than measuring an error.
Sampling is distinct from recasting. The calendar keeps its structure; only some periods are retained, and those carry a weight scaling them back up to a full year. A sample must keep at least two members at every level below the year, because the surviving level columns are what define the leaf timeslices. Dropping a level entirely renames the leaves, the model’s data silently stops matching, and the result solves to zero.
Processes and variants
A process may be described once or as a family of variants: vintages
differing by build year, clusters differing by resource class, loss
tranche or price step. expand_variants() turns the
declaration into separate processes and combine_vintages()
performs the reverse for vintages.
What a cluster means differs by class, and this governs what may be merged. On a technology or a storage it is a parallel sub-process; on a trade, a tranche of a piecewise-linear loss curve; on a supply, import or export, a step of a price curve; on a demand, a sub-regional share. Tranche efficiencies are derived from tranche shares, so the two cannot be merged independently.
Reducing size without changing resolution
interpolate_model(fold = , sparse = , prune = ),
trim_parameters() and drop_default_values()
make a model smaller without changing what it represents. These are
representation choices rather than resolution choices and they do not
alter the answer. model_size() reports what they
achieved.
Direction
The balance layer aggregates and never disaggregates. Fine commodity totals are summed into coarse ones; a coarse quantity is never split across finer cells. This is deliberate, and it is why an object may be declared coarser than the commodities it consumes but never coarser than those it produces or stores.
In consequence, disaggregation can only mean building a finer model
before interpolation. No operation takes a solved coarse result and
distributes it over finer regions or slices, and none is planned. Where
a region subset would split a group that some object is declared over,
subset_model_regions() stops rather than guessing: an
unchanged parent resource overstates a shared pool, a proportional share
understates an indivisible one, and there is no safe default. The error
names the three ways forward — extend the region set, aggregate to the
coarser level first, or declare the objects on the finer regions.
What is not supported
Several operations are refused rather than approximated, each for a stated reason.
- Aggregating a weather stream carrying a non-linear transform: the mean of a transform is not the transform of the mean.
- Aggregating weather between two levels when neither is the finest, which needs a level-aware crosswalk that does not exist.
- Aggregating trade routes that carry weather factors.
- Grouping corridors carrying reactance or resistance above the floor: parallel AC circuits divide flow by impedance, not by optimisation.
- Cross-cutting groupings, where a region belongs to two parallel hierarchies that do not nest.
- Region grouping for demand, weather and commodity objects, none of which is a process with archetypes.