Oxford grid-battery real-data example

This opt-in example applies the bcla workflow to the University of Oxford’s energy-trading battery degradation dataset. The source contains one year of operation for six 16 Ah lithium-ion cells following three grid-trading control strategies, with per-cell current profiles and 13 capacity checks per cell.

The repository does not bundle, mirror, or relicense the dataset. The example downloads original files from the Oxford Research Archive at runtime. Each download is capped at 32 MiB before decoding; the largest official profile was 8.87 MiB during the 2026-08-31 audit. A larger response is rejected rather than read into memory without a bound.

• Official dataset record: https://ora.ox.ac.uk/objects/uuid:9aae61af-2949-49f1-8ad5-6aea448979e5
• Dataset DOI: 10.5287/bodleian:gJPdDzvP4
• Data license: ODC Open Database License (ODbL) 1.0; individual contents are covered by the Database Contents License (DbCL) 1.0.

Run it

From an installed source checkout:

python examples/oxford_energy_trading.py --cell BMR_cell1

To save a diagnostic figure or disable the residual bootstrap:

python examples/oxford_energy_trading.py \
  --cell BMR_cell1 \
  --bootstrap-samples 0 \
  --output oxford_bmr_cell1.png

The six supported cell IDs are BMP_cell1, BMP_cell2, BMR_cell1, BMR_cell2, SPM_cell1, and SPM_cell2.

Converting measured current to EFC

The capacity files report profile_time_s and capacity_Ah; they do not report a laboratory cycle index. Relabeling elapsed time as cycle count would be misleading. The example instead aligns each capacity check with its cell’s measured time_s,current_A profile and estimates cumulative discharge equivalent full cycles:

EFC(t) = integral(max(I(t), 0) dt) / (3600 * Q_nominal)

Positive current is discharge in the source data, negative current is charge, and Q_nominal is the stated 16 Ah cell capacity. Charging samples are clipped to zero before trapezoidal integration. The result is discharge-throughput EFC, not elapsed days, commanded cycles, or a count of laboratory charge/discharge procedures.

The official files contain a small number of duplicate or reversed profile-time rows. The example makes its repair observable: it stable-sorts profile time, averages current at identical timestamps, and prints the number of affected transitions and duplicate samples. Each final capacity check also trails the last near-zero-current sample by less than 1,004 seconds. The example caps that small tail to the last observation only when the gap is at most 1,800 seconds and the last absolute measured current is at most 1 mA; it prints the gap. Only that single final capacity check may be capped. Anything outside those conditions is rejected.

Saved figures relabel the horizontal axis as Measured Discharge EFC; they do not present this throughput estimate as a laboratory cycle count. Solid lines cover the observed window and dashed lines show only the bounded projection to EOL, or to the three-times-observed limit when EOL is unsupported. The console output states the chronological training/holdout sizes and reports bounded EOL, RUL, and bootstrap censoring diagnostics explicitly.

Live source audit

The following values were reproduced from the original ORA files on 2026-08-31 using the same canonicalization and integration code:

Cell	Capacity checks	Measured discharge EFC	Final retention
BMP cell 1	13	412.37	97.53%
BMP cell 2	13	410.12	97.56%
BMR cell 1	13	1558.15	85.97%
BMR cell 2	13	1557.65	85.15%
SPM cell 1	13	689.23	98.35%
SPM cell 2	13	689.93	98.23%

The measured EFC is lower than an estimate from commanded schedules because the tester enforced voltage limits and entered constant-voltage operation when necessary. That is why the measured per-cell current files are primary here.

Interpretation limits

• EFC integration is approximate because logging was irregular and samples were not guaranteed exactly at current transitions.
• Thirteen monthly capacity checks provide a compact demonstration, not broad forecast validation or independent-cell cross-validation.
• The 20% chronological holdout contains three observations. Report that small sample alongside every validation score.
• Bounded EOL can be unsupported, and bootstrap replicates can be censored. The example prints those outcomes instead of dropping them.
• The empirical fits do not model electrochemistry, cell safety, dispatch, or future protocol changes.

Citation

When using the data, follow the archive’s attribution and ODbL/DbCL terms and cite the dataset:

Reniers, J. M., Mulder, G., and Howey, D. A. (2020). Oxford energy trading battery degradation dataset. University of Oxford.

The source readme also requests citation of the related work, including:

Reniers, J. M., Mulder, G., and Howey, D. A. “Improving optimal control of grid-connected lithium-ion batteries through more accurate battery and degradation modelling.” Journal of Power Sources 379 (2018), 91-102. 10.1016/j.jpowsour.2018.01.004