Archived v5 diagnostic. These results describe the September 14, 2026 model and do not validate the current site score or predict which new projects will succeed. Read the model review.
How did observed battery sites score in v5?
In 59 matched comparisons, the median observed-minus-alternative score difference is -0.2 points. Top-quintile enrichment is 1.0.
This is a descriptive diagnostic. No weights were fitted to these projects. Existing batteries do not receive points simply because they appear in the inventory. Higher agreement with observed locations alone does not prove that a score predicts successful development.
In this sample, observed locations and matched alternatives receive broadly similar scores. The recorded v5 score does not demonstrate strong separation of real-world development locations.
The comparison
Every location is evaluated as a 100 MW, four-hour standalone battery, target 2031, minimum 115 kV, 25 km study radius. The assumptions are held constant; this is not a reconstruction of each project's actual sizing, contract or hybrid configuration.
EIA inventory: 2026-07. CAISO queue: 2026-09-10. Hourly electricity prices: 2025. Environmental snapshot: 2827bcdb6114.
The primary inventory cohort requires at least 20 MW of battery capacity in the relevant maturity category. Out-of-service units are excluded. Active queue hybrids are excluded from the primary queue-only cohort because their published project MW does not identify battery MW; EIA solar-colocated sites have separately reported battery MW and remain flagged in the data.
After grouping colocated records, this run evaluated 174 observed groups and 600 alternative locations. It matched 59 pairs; 115 observed groups had no unused alternative satisfying all comparison criteria. The source inventory contains 360 plants; 148 active/planned plants fell below the size threshold and 6 were entirely out of service.
Results by development status
| Observed cohort | Spatial groups | Matched pairs | Observed median | Alternative median | Median paired difference | Top-quintile enrichment |
|---|---|---|---|---|---|---|
| operating | 99 | 27 | 55.2 | 54.0 | 0.3 | 1.3 |
| construction | 20 | 5 | 55.7 | 53.6 | 3.1 | 1.0 |
| planned | 23 | 8 | 54.4 | 54.4 | -0.1 | 1.0 |
| queue-only | 32 | 19 | 55.0 | 55.0 | -1.1 | 1.3 |
Archived v4 and v5 methods
| Measure | Archived v4 | Archived v5 |
|---|---|---|
| Method version | 2026-09-unified-screening-v4 | 2026-09-battery-poi-v5 |
| Matched observed median | 55.8 | 55.2 |
| Matched alternative median | 57.3 | 54.2 |
| Median paired difference | -1.2 | -0.2 |
| Observed above alternative (%) | 46.6 | 49.1 |
| Top-quintile enrichment | 1.1 | 1.0 |
The same matched locations and fixed scenario are used for both archived methods. V5 leaves unmatched POI queue evidence unknown and uses candidate POI price-node associations where supported. V4 penalized geographic queue volume and always selected the geographically nearest price node. Missing-evidence normalization can itself change scores; an increase is not evidence of an economic improvement. Neither column represents the current application.
What drives the difference?
| Component | Observed median | Alternative median | Median paired difference | Known pairs |
|---|---|---|---|---|
| grid | 99.0 | 99.9 | -0.8 | 59 |
| resource | 20.1 | 19.0 | 0.0 | 59 |
| development | 65.3 | 59.7 | 0.0 | 59 |
| market | 13.8 | 12.8 | 0.0 | 59 |
| queue | 50.0 | 50.0 | 0.0 | 2 |
These are component scores before weighting. A positive difference means the observed sites score higher on that component. Unavailable components are excluded from the component comparison and their known-pair count is shown.
Evidence coverage
| Observed cohort | Observed full coverage (%) | Alternative full coverage (%) | Observed queue known (%) | Alternative queue known (%) |
|---|---|---|---|---|
| operating | 80.9 | 80.9 | 25.9 | 7.4 |
| construction | 78.9 | 80.9 | 20.0 | 20.0 |
| planned | 80.9 | 80.9 | 12.5 | 0.0 |
| queue-only | 94.9 | 80.9 | 63.2 | 5.3 |
Matching uses coverage of the non-queue components only, normalized to 100%. It does not select on whether a queue association is known. The table reports full production-score coverage separately because unknown queue evidence can affect score normalization.
How the comparisons are selected
Observed records are grouped by their nearest qualifying substation within 5 km, otherwise by a 0.05° geographic cell. Each group contributes one anchor: most mature qualifying status, then largest capacity, then stable ID. Groups can contain several statuses; every member is retained in the downloadable data.
Alternatives are mapped California substations without a mapped EIA battery or active battery/hybrid queue location within 5 km. They must match the observed site's voltage band, lie within 200 km, and differ by no more than 2 km in substation proximity, 3 km in line proximity, five percentage points in non-queue evidence coverage and eight points in grid score. Matching never uses the overall, market or queue score. Controls are assigned without replacement; unmatched cases remain visible.
The observed project's own queue record is excluded where a conservative link exists. A unique name-token match of at least 0.8 and distance of at most 10 km is required for an EIA-to-queue link. 29 primary inventory sites were linked; unresolved identities can leave residual self-influence. Queue-only anchors exclude their own queue ID, while other developers at the same POI remain in the evidence.
What this cannot establish yet
- 2025 prices and 2026 evidence cannot validate predictions made before older projects were built.
- An operating battery shows past development feasibility. A proposed project shows interest. Neither establishes today's best remaining site.
- Alternative grid locations are not confirmed available parcels or failed projects. Nearby substations may share electrical constraints, so broader spatial correlation remains.
- Contracts, resource-adequacy revenue, verified charging/export rights, hybrid coupling and detailed connection costs are not reconstructed here.
- The matching criteria select a subset of the inventory. Review unmatched cases and evidence coverage alongside the matched results.
A future predictive test should freeze dated inputs, observe subsequent outcomes and reserve whole POIs or regions before adjusting weights. This report establishes an auditable comparison for that work.
Data and reproducibility
Run node siting/scripts/validate-battery-siting.mjs against the bundled files. This report used zero external requests. The JSON includes the scenario, per-location results, component scores, matches, exclusions and SHA-256 hashes for all consumed datasets.
The previous evaluator is preserved in the v4 baseline archive. Extract it and pass its model path with --baseline=path/to/baseline-v4/model.js to reproduce the before-and-after comparison using the same local data.