Chile curtailment, nodal prices and equivalent storage#

This study implements the core screening methods in Acosta et al., Journal of Energy Storage 183 (2027), 125203, available online 22 October 2026: DOI 10.1016/j.est.2026.125203. It reproduces published aggregate findings, rather than independently reprocessing the authors’ raw CEN dataset. The supplied article states that its data are available on request. The PDF is not redistributed in this repository.

Run the complete offline example#

From the repository root:

uv sync --locked --all-packages
uv run python -m energy_analysis.chile_curtailment --output artifacts/chile-curtailment
uv run python examples/analysis/plot_chile_curtailment_storage.py

The CLI writes annual results, duration scenarios, Scenario A/B, 80 sensitivity cases, and 35 numerical comparisons as CSV and Parquet, plus a manifest with the source fixture hash, assumptions and output hashes. No API key or network access is needed after installation. The gallery generates two figures.

Comparison

Recomputed from the published inputs

Paper

2022 curtailed energy

1471.0 GWh

1471.0 GWh

2025 curtailed energy

6044.1 GWh

6044.1 GWh

2025 curtailed / generated

18.492%

18.49%

2025 PV curtailed / generated

21.592%

21.59%

2025 wind curtailed / generated

13.282%

13.28%

2025 4-hour effective EESC

3539.524 MW

3539.5 MW, Table 8

Effective gap versus 1714 MW installed

1825.524 MW

about 1825 MW

Block B effective EESC

2900.851 MW

2900.8 MW

Reduction relative to Scenario A

18.044%

18.0%

Values retain the precision of Tables 2, 3, 7 and 8. Table 8’s Scenario B differs by 0.051 MW from Eq. (2) applied to its rounded energy inputs; the audit exposes that residual and uses a stated 0.1 MW tolerance. It does not tune the inputs to force identical final digits. Table 7 uses a 1 MW tolerance; percentages use 0.01 percentage points. 34 of 35 comparisons agree within those tolerances. The remaining comparison is preserved as a source discrepancy: the 2024 table inputs give 19.0009%, while Table 3 reports 18.99%. Even using the rounded 5642 GWh in Table 3 gives approximately 19.00%. We retain both values and the failed agreement flag rather than widening the tolerance or changing the source. The regression suite requires this known disagreement to remain visible. The fixture records the supplied PDF’s SHA-256 and table/page locations. Source figures are literature aggregates, not measurements made by this project.

Definitions that change the interpretation#

  1. Curtailment denominator. Table 3 uses curtailed energy / reported generation. The alternative curtailed / (generated + curtailed) is about 15.61% in 2025. Both columns are exposed; they must not be interchanged.

  2. Equation 2. EESC_MW = curtailed_MWh * DoD * efficiency / (days * hours). The implementation preserves multiplication by DoD and efficiency. Lower efficiency lowers this effective-power indicator. It does not lower the nameplate battery capacity needed to capture a fixed energy volume.

  3. Scenario B. Only energy in local interval starts 08:00 <= hour < 18:00 is included. This follows the prose definition of Block B despite the ambiguous inclusive upper bound in the printed Eq. (3). Dropping the other 1090.6 GWh changes the energy scope; it does not prove that a smaller battery captures all annual curtailment. The code reports both (A-B)/A (about 18%) and (A-B)/B (about 22%); those denominators answer different questions.

  4. Calendar. Offline reproduction uses the paper’s fixed 365 days for every year, including leap-year 2024. Other studies can pass their actual day count.

  5. Price separation. Eq. (1) uses abs(price_i-price_j) >= 10 USD/MWh. Equality is flagged, negative prices are allowed, and the result is named price_separation_flag. This is a proxy, not a DC/AC power-flow diagnosis.

  6. Joint stress. Table 1 assigns I=low/low, II=low PV/high wind, III=high PV/low wind, IV=high/high. P50 uses all paired observed hours and linear interpolation. Equality is high, so a zero median can label zero curtailment as high. The thresholds are reported for inspection; they are not silently replaced with positive-only medians.

There is no chronological dispatch, SOC tracking, degradation, siting, transmission model, revenue optimization or investment recommendation here. The reported installed MW are not an assertion that every installed BESS has four hours of usable energy or can serve the constrained nodes. The indicator is not an established upper/lower bound on a feasible battery design.

Connect real CEN exports#

The connector is energy_data_retrieval.sources.cen.curtailment. It supports preserved long-form CSV and observed public CSV download URLs on CEN hosts, using the existing bounded HTTP transport. No unverified curtailment or nodal price API endpoint is invented. CEN’s source categories are the ERNC reduction report, SEN generation report and preliminary real marginal costs by busbar. The existing sources.cen.client.CENClient remains the authorized generation API route. Excel/Qlik files must first be exported to long-form CSV; keep the original source file and its interpretation evidence.

from pathlib import Path
from energy_data_retrieval.sources.cen.curtailment import (
    CENHourlySchema,
    import_cen_hourly_csv,
    fetch_cen_hourly_csv,
)

schema = CENHourlySchema(
    quantity="curtailment",
    timestamp_column="timestamp",
    entity_column="plant_id",
    value_column="curtailed_mwh",
    unit="MWh",
    timezone=None,
    technology_column="technology",
    technology_map=(("Solar", "pv"), ("Eolica", "wind")),
)
# observed_source_url must be the actual public CEN source URL you inspected.
export = import_cen_hourly_csv(Path("curtailment.csv"), schema, source_url=observed_source_url)
# With a configured HTTPClient: fetch_cen_hourly_csv(http, observed_csv_url, schema)

The source mapping is deliberately explicit: source column names vary across exports. timezone=None requires offsets in the source timestamps. A declared source zone permits unambiguous naive labels; ambiguous/nonexistent DST labels fail. A repeated hour with explicit different UTC offsets remains two records. Declare whether labels refer to the start or end of the interval. MW means hourly average power for exactly one hour; GWh converts to MWh. Subhourly, daily and instantaneous power records are not supported by this adapter.

The reader rejects missing/nonfinite values, negative energy, unknown technology labels and duplicate entity-hours. It retains identifiers such as 001. It does not infer that all plants or hours are present. A source URL and input hash record provenance; a local CSV is not independently authenticated as CEN data.

Hourly analysis configuration#

Provide a JSON object with curtailment, optional prices, expected_entities, optional corridors, and optional timezone (default America/Santiago). Each export object contains path (relative to this JSON file), source_url (the observed CEN URL) and schema (the CENHourlySchema fields above). For prices, set quantity="marginal_cost", unit="USD/MWh", and map the entity column to the busbar identity. Price schemas do not need a technology map.

expected_entities maps pv and wind to lists of verified plant identities. Every observed hour must include that full roster; changing rosters require separate time slices. corridors is a list of explicit node pairs, for example [["Cardones", "Pan de Azucar"]] only if those are the actual source IDs. Use verified source identifiers rather than matching similar names.

uv run python -m energy_analysis.chile_curtailment \
  --output artifacts/chile-curtailment-observed --hourly-config local-config.json

Hourly output includes P50 quadrants, local blocks/seasons, seasonal shares of all quadrant-IV hours, and sample coverage. Whole missing hours remain missing and are counted over the observed first-to-last span. Price pairs use common observations and report the number of unpaired hours. Seasonal shares describe the supplied sample, not necessarily a complete year. No extrapolated annual EESC is computed from a partial hourly export.

The paper reports spring/summer shares of high-stress events and statistical tests on monthly data. Those are not independently reproduced: the raw hourly/monthly series are unavailable, and the paper describes BCa bootstrap in section 3.2.4 but percentile bootstrap in Table 4. The implementation does not manufacture samples or select a statistical variant to match reported p-values.

Implementation map#

Layer

File/module

Purpose

Methods

energy_methods.curtailment_storage

Ratios, EESC A/B, sensitivity, P50 quadrants, price screens

Connectors

energy_data_retrieval.sources.cen.curtailment

Explicit CEN CSV normalization and bounded download

Analysis

energy_analysis.chile_curtailment

Reproduction audit, observed hourly diagnostics, CSV/Parquet CLI

Data

analysis/src/energy_analysis/data/chile_curtailment_2027.json

Attributed published aggregates and source hash

Example

examples/analysis/plot_chile_curtailment_storage.py

Offline runnable gallery figures

Original implementation by ASVA Consulting, AI-assisted and reviewed against the supplied literature. Existing repository licensing applies to the code; no license to the publisher’s PDF or underlying operator dataset is asserted.