Skip to content

Generator Clustering

PowerGenome groups real-world power plants into representative clusters to keep capacity expansion models tractable. Rather than modeling each unit individually (which could mean thousands of rows), generators with similar characteristics are merged into a smaller number of clusters per region and technology type.

Why clustering is needed

Detailed generator inventories from EIA data can contain thousands of individual units. Including each unit as a separate resource in GenX would make the optimization problem very large without meaningfully improving accuracy, because units of the same vintage and technology behave nearly identically.

Clustering reduces these to a manageable number of representative resources while preserving the key features the model needs: capacity, heat rate, cost, and retirement eligibility.

How existing generators are clustered

Step 1: Filter and group

Generators in the model regions (after region aggregation) are grouped by (model region, technology) pairs. Each group is clustered independently. Retirement status determines which units contribute capacity, not cluster membership.

Step 2: K-means clustering

Within each region–technology group, k-means clustering partitions units into k clusters where k is set by num_clusters (or alt_num_clusters for per-region overrides).

Clustering features (by default):

  • heat_rate_mmbtu_mwh — thermal efficiency
  • fom_per_mwyr — fixed O&M cost

Features are standardized (mean 0, unit variance) before clustering so that neither dominates the distance metric purely because of scale.

If a region–technology group has fewer plants than the requested number of clusters, the number of clusters is automatically reduced to match the number of plants.

Step 3: Cluster aggregation

For each cluster, representative values are computed:

Attribute Aggregation method
capacity_mw Sum of operating units in cluster
heat_rate_mmbtu_mwh Capacity-weighted average
fom_per_mwyr Capacity-weighted average
vom_per_mwh Capacity-weighted average
minimum_load_mw Capacity-weighted average

Step 4: Resource naming

Each cluster is assigned a resource label:

<region>_<technology_snake_case>_<cluster_number>

For example: CA_N_natural_gas_fired_combined_cycle_1

This name is used as the resource identifier throughout the output files.

Missing existing-resource availability profiles

Existing hydro, wind, solar, and other profile-backed resources need usable hourly availability data. PowerGenome checks missing resource groups, missing regional metadata, missing sites, and incomplete time series against a combined regional fallback budget before assigning constant 1.0 availability.

PowerGenome permits a small, combined regional fallback fleet: by default, existing_resource_profile_fallback_max_demand_fraction: 0.005 allows affected operating capacity up to 0.5% of mean regional demand before DG subtraction. The denominator includes regional aggregation and supplemental/flexible demand, uses unrounded hourly values for the selected scenario and years, and precedes time reduction. Means across weather years are weighted by modeled hours.

At or below the limit, a warning identifies the affected resources and PowerGenome retains their capacity, tags, and retirement attributes with availability 1.0 in all modeled hours. Above it, an actionable error stops generation of model inputs. Partial profile coverage charges the entire failing existing-generator cluster; resource-group metadata capacity is never used as the affected capacity. All affected technologies and clusters share the allowance in their model region. Valid profiles remain unchanged, and retired/zero-capacity clusters consume no allowance.

For example, SD (p32 + p38) with 1584.86 MW mean demand has a 7.92 MW allowance: 4 MW of missing run-of-river profiles warns and falls back. Modeling p32 alone with 206.28 MW mean demand gives only 1.03 MW, so the same 4 MW errors. If those units retire in 2030, they contribute zero operating capacity at model year 2030, even if the planning period starts in 2026.

Set the fraction to 0 for strict mode; values must be finite numbers in [0, 1], not booleans. Positive-capacity fallback requires valid demand even with --no-load; complete profiles add no demand requirement. Explicitly profile-free resource groups (group.profiles is None, such as thermal/storage groups) remain exempt, as do unsupported dispatchable technologies. New-build behavior is unchanged. Invalid schemas, unreadable files, ambiguous group matches, and invalid weather-year requests remain errors rather than fallback conditions.

This tolerance is a modeling choice, not an industry standard or reliability guarantee. Constant availability can overstate usable generation. Its full-output energy is bounded by the chosen fraction of gross demand over the same hours, but that bound says nothing about hourly reliability or dispatch. Repair metadata/profile coverage before considering a larger tolerance.

Controlling the number of clusters

Default cluster count

num_clusters: 1  # Default: merge all plants of a type into one cluster

A value of 1 means all generators of the same type in a region become one representative unit. This is the most aggressive reduction and is appropriate when you need a compact model.

Per-region overrides

Use alt_num_clusters to set different cluster counts for specific combinations:

alt_num_clusters:
  CA_N:
    Natural Gas Fired Combined Cycle: 3   # Three CCGT clusters in CA_N
    Conventional Steam Coal: 1             # All coal in one cluster
  TX:
    Natural Gas Fired Combined Cycle: 5

Setting num_clusters to 0 or omitting a region/tech combination excludes those generators entirely. Technology names here are matched against the raw data values; they are not the ATB-style names.

Individual units (no clustering)

To preserve every generator as an individual resource, set num_clusters equal to the number of plants (or use a large number — PowerGenome reduces it to the number of plants automatically):

num_clusters: 999  # Each plant becomes its own resource

Technology grouping

tech_groups merges similar technologies into a single group before clustering, so they share a resource label:

tech_groups:
  Landfill Gas:
    - Landfill Gas
    - Municipal Solid Waste
    - Other Gases

This is useful for small fuel types where separating them would create many tiny resources.

Retirement filtering

Set scheduled retirements in the generation table's retirement_year column. A blank value or missing column means no planned retirement.

A unit contributes its full capacity if it is operating at the planning period's end (model_year): operating_year must be on or before that year, and retirement_year must be later or unspecified. Units that come online during a period are included without prorating their capacity.

Myopic multi-period models

Use the same generation data across the periods of a case to keep cluster membership consistent. PowerGenome records scheduled retirements in later periods as minimum retirement requirements for each resource (Min_Retired_* columns).

Viewing cluster assignments

When extra_outputs_path is configured (or by default in the extra_outputs sub-folder of each case), PowerGenome writes a CSV file showing which real plant belonged to which cluster. This is useful for debugging unexpected groupings.