33 feeders with a usable real aggregate shape| statistic | value |
|---|---|
| count | 33.0 |
| mean | 769.1818181818181 |
| std | 640.6141806181713 |
| min | 1.0 |
| 25% | 125.0 |
| 50% | 767.0 |
| 75% | 1126.0 |
| max | 2580.0 |
Chapter 5: Feeder-Level Clustering at Scale
Chapter 3 clustered customers. Ledva and Mathieu and Li et al. ask the same question one level up: do whole feeders fall into a handful of aggregate archetypes too, residential-heavy, commercial-heavy, mixed, useful for a Distribution System Operator (DSO) planning investment across many feeders at once, not just one at a time [1], [2]. AusNet’s own single 31-customer feeder cannot answer this, there is only one of it. This chapter uses SMART-DS, National Renewable Energy Laboratory (NREL)’s real, validated synthetic distribution testbed for the Austin, TX region, 33 real Low Voltage (LV) feeders across 8 substations, each with its own real per-building demand profiles, to ask the question on a real population of feeders rather than one.
Getting the SMART-DS data
Each feeder’s network model uses 200 to 350 real customer profiles from a shared, region-wide pool. Fetching every profile for all 33 feeders would mean thousands of small downloads, even though this chapter only needs each feeder’s overall load shape, not every individual customer. Instead, the fetch script samples a real, kW-stratified subset per feeder from the same shared pool:
uv run python scripts/fetch_part4_smartds_data.pyTwo small helpers do the work: feeder_load_profiles() parses a feeder’s own Loads.dss into (customer name, kW, profile id) triples, and feeder_aggregate_shape() sums whichever sampled profile CSVs actually landed on disk into one real, kW-weighted aggregate demand curve per feeder.
Since these are common, real patterns found throughout the region, and not unique to each customer, even a small sample from each feeder represents a much larger part of that feeder’s customer base than the sample size alone suggests: samples drawn for other feeders often include patterns this feeder’s own customers also happen to use.
Coverage varies. Some feeders have thin coverage, only a handful of real customers whose sampled shape happened to match, worth keeping in mind: their own cluster membership rests on less real evidence than a feeder with hundreds of matched customers (median 767, ranging 1 to 2,580), the same “how many meters do you need” caution Part 4 Chapter 3’s own sensitivity section raised for customer-level phase identification.
Clustering the feeders themselves
Clustering at the feeder level gives much clearer separation than at the customer level: a silhouette score of 0.67 at k=2, against 0.23 for individual customers. This is expected, not a contradiction. Averaging at the feeder level smooths out the individual noise that caused overlap at the customer level in the first place. Unlike the customer-level case, k=2 actually is the validity-index optimum here, not a compromise past it.
Two distinct feeder archetypes emerge. The first shows a sharp evening peak and much lower daytime demand, consistent with a residential-heavy mix. The second holds a sustained plateau through the middle of the day before a noisier, more irregular evening, consistent with a commercial- heavy mix whose demand does not track the residential dinner-and-bedtime rhythm Chapter 3 was clustering on. Exactly the “different customer mix” a DSO would want to plan for differently, now visible at the feeder level, not just guessed at.
Where this leads
Two real feeder archetypes, the same shape-based recipe Chapter 3 already settled on, applied one level up rather than re-derived for a different granularity. This is the first, smaller piece of a larger question this chapter does not yet fully answer: the same “cluster whole feeders” question asked again at the substation rather than the feeder, and on real networks larger than SMART-DS’s own 33, AusNet’s own SMR8, KLO14, HPK11, and CRE21 networks among them. li2015substation’s own substation-clustering method, cited here for the same feeder-clustering motivation ledva2020feederdemand already provides, is a real, waiting anchor for that larger question, not yet built.