Research benchmark · Research benchmark

VoltRL

Benchmark causal battery policies without pretending the benchmark is a live market system.

VoltRL fixes information timing, state representation, battery constraints, degradation assumptions, holdout separation, and result provenance so policy comparisons remain auditable.

Open repository Read documentation Run or reproduce Inspect tests and evidence

Engineering question

How do finite-state policies compare when information timing, holdout separation, battery constraints, degradation assumptions, and diagnostic bounds are explicit?

Audit finite-state battery-arbitrage policies under explicit synthetic and historical information protocols without presenting the benchmark as live bidding.

System and evidence boundary

State, information, action, and settlement timing stay separate.

The benchmark compares price-only and price-plus-hour protocols while keeping perfect foresight outside the causal policy set.

Finite-state control loop

01Price stateDiscretized sequential observation
02SOC state0–500 MWh grid
03Optional UTC hourInformation-protocol ablation
04Charge / idle / dischargeFeasible one-hour action
05Reward + aging costSettlement less explicit stress proxy
06Next stateTransition and stored-energy update

Chronological evidence protocol

0170% train / developmentFit transitions, forecasts, and model resolution
0224-hour schedule fixedBefore each UTC delivery day
0330% untouched holdoutNo same-day realized-price replanning
04Realized settlementPrice-taking diagnostic evaluation
05Perfect foresightUpper bound only—not a causal policy

Artifact-integrity chain

01OPSD packageVersion + DOI + local SHA-256
02Experiment configurationManifest + pinned dependencies
03CSV result bundleSeed, market, sensitivity, diagnostics
04PNG + PDF figuresSix release figure pairs
05Independent audit13 release tests; post-release audit utilities disclosed separately

Verified results

Figures keep scenario, units, source, and evidence class attached.

Every visual below points to its source data or retained source format and to the release- or commit-specific implementation behind it.

VoltRL release figure showing synthetic and historical price data, chronological partitions, and model-resolution selection diagnostics — Synthetic + historical · release artifact
Data partitions and training-only model selection

Copied from v1.1.0 results_revision2. Historical panels use OPSD DK1/DK2 day-ahead data; synthetic panels use 30 independently regenerated seeds.

Provenance
v1.1.0 · 70fa9609a1a1 · verified 2026-07-28

Reproduce
`python scripts/generate-open-engineering-figures.py --sources-root <audit-root>`

Download source PNG Download source PDF Inspect repository source
Open full-size figure

VoltRL comparative results figure with policy-performance panels for synthetic state protocols and historical DK1 and DK2 schedules — Synthetic + historical · release artifact
Policy comparison under explicit protocols

The price-only and price-plus-hour comparison is separate from historical block scheduling. Perfect foresight is a diagnostic bound, never an available causal policy.

Provenance
v1.1.0 · 70fa9609a1a1 · verified 2026-07-28

Reproduce
`python scripts/generate-open-engineering-figures.py --sources-root <audit-root>`

Download source PNG Download source PDF Inspect repository source
Open full-size figure

VoltRL release figure comparing historical forecast diagnostics and degradation-model sensitivity — Historical + sensitivity · release artifact
Forecast and degradation sensitivity

The released figure combines day-ahead forecast diagnostics with explicit no-cost, linear, and nonlinear degradation-model sensitivity.

Provenance
v1.1.0 · 70fa9609a1a1 · verified 2026-07-28

Reproduce
`python scripts/generate-open-engineering-figures.py --sources-root <audit-root>`

Download source PNG Download source PDF Inspect repository source
Open full-size figure

Validation and evidence

What passed, what was demonstrated, and what remains unvalidated.

Thirteen deterministic tests passed; the release contains a manifest, CSV results, and six PNG/PDF figure pairs.

Release tests

13 deterministic tests passed for SOC feasibility, information timing, block commitment, nonlinear degradation normalization, solver behavior, and oracle dominance.

Result bundle

v1.1.0 includes seed metrics, bootstrap comparisons, historical schedules, forecast diagnostics, sensitivities, solver checks, a manifest, and six PNG/PDF figure pairs.

Data provenance

The required OPSD 2020-10-06 filename, DOI, official package URL, two market columns, and SHA-256 are recorded; the 124 MB source CSV is not committed.

Reproduction limitation

The full benchmark was not rerun during this site audit because the large OPSD input was absent. Release tests passed and the committed result bundle was inspected byte-for-byte.

Shortest verified path

Pin the audited source before running it.

git clone https://github.com/mohammadrezwankhan/voltrl.git
cd voltrl
git checkout v1.1.0
python -m pip install -r requirements.txt
# Download and checksum the official OPSD 2020-10-06 CSV first.
python voltrl_benchmark.py --opsd-csv data/opsd_time_series_60min_singleindex_2020-10-06.csv --output-dir results_revision2 --synthetic-seeds 30 --synthetic-hours 17520 --candidates 4,6,8,10,12,16,20,24,32,40,48
python -m unittest discover -s tests -v

Verification environment: Python 3.12.13 with the pinned v1.1.0 dependency stack on Windows 11. Distinguish running the scientific model from rebuilding site figures (which uses scripts/generate-open-engineering-figures.py in the portfolio website repository). Commands are concise entry points; repository documentation remains authoritative for setup details.

Documentation map

Go directly to the useful evidence.

Engineering boundary

The artifact cannot qualify the decision by itself.

VoltRL is a research benchmark, not a deployment, live bidding engine, commercial optimizer, or market-access system.
Historical schedules assume accepted fixed quantities, price-taking settlement, and one block committed before each UTC delivery day; fees, bids, imbalance, market impact, and operational dispatch are omitted.
The battery is a finite energy-and-power abstraction with stated efficiencies and an explicit degradation proxy, not an electrochemical, thermal, warranty, or safety model.
Perfect foresight is only a diagnostic upper bound. It must never be described as achievable causal policy performance.

Related professional context

Follow the evidence into adjacent work.

Use the pinned source, inspect the evidence, and keep project authority with a qualified human reviewer. Open v1.1.0 Return to open engineering

Evidence summary

Reviewed 2026-11-12; evidence records retain their own scope and dates.

Documentation snapshot: 2026-10-06 · a2f7eed · a2f7eed0886a5a4844e8db2764c0bfd88615d1d8
Portfolio source record: 13 results · executed-upstream · 70fa960 · 70fa9609a1a122055f50f7897720eefc97cbb3e0
Current release: v1.1.0 · published 2026-07-16 · tag target 70fa9609a1a122055f50f7897720eefc97cbb3e0
Current verification: Successful hosted run · 2026-11-09 · target b244bc36b512610092064b5f903ed5f08466dc7b
Repository head at 2026-10-17 inventory: a2f7eed0886a5a4844e8db2764c0bfd88615d1d8 · inventory snapshot only