# Similar summaries, different recorded paths

**Two pairs of BESS records — and a question about what a summary leaves out.**

A battery asset owner needs to know what the asset can deliver under the next request. A compact condition summary is useful, but the path by which a battery charges or discharges may contain differences that the summary does not describe.

These examples make one limited point visible: **close starting SOC records, close duration and close supported Ah do not specify the same recorded current and voltage path.** Whether preserving those paths improves a particular operational decision remains to be tested. SOH is not compared in this case.

## Observed

The figures retain all nine racks across each complete retained period. Their dots are records, not a continuous or synchronized picture of the physical battery. The examples were selected after inspection; they are not a representative sample.

| Pair | Maximum first-SOC difference, same rack | Container-duration relative difference | Maximum supported-Ah relative difference, same rack |
|---|---:|---:|---:|
| Discharge #130 / #198 | 1 percentage point | 0.87% | 1.63% |
| Charge #222 / #276 | 1 percentage point | 0.70% | 0.66% |

Relative difference is `2 × |a − b| / (a + b)`. “Supported Ah” is only the integral over specified recorded-current intervals; it is not full transferred charge or capacity. The underlying query allowed at most 2 percentage points in first SOC and 5% in each of the other two comparisons.

### Discharge: the amount does not specify its timing

Record #130 (18 December 2024) lasts 128.25 minutes; #198 (17 January 2025) lasts 129.37 minutes. Yet their current paths differ visibly. In #130, an earlier lower-current plateau precedes an increase. In #198, the higher-current portion arrives earlier, with current stepping down later. During that later part, the retained voltage records turn upward.

![All nine racks in discharge records 130 and 198: native current, rack voltage, reported SOC and partial supported Ah, on shared channel scales.](discharge.png)

**Read the current row before interpreting the voltage row.** The realized current histories are different. This is not evidence that an identical input produced a different battery response, and it does not isolate aging, a material mechanism or a controller decision.

### Charge: keep the members and their timestamps

Records #222 (30 January 2025) and #276 (17 February 2025) also meet the three proximity rules. The full plots preserve the late current separation among racks and the member-specific reported SOC records rather than reducing either period to a single endpoint.

![All nine racks in charge records 222 and 276: native current, rack voltage, reported SOC and partial supported Ah, on shared channel scales.](charge.png)

One concrete example is R3: its first-to-last reported SOC is 4→95% in #222 and 5→100% in #276, while its supported charge-in subtotals are 222.32 and 222.97 Ah. However, the last SOC record is 38.5 seconds before the container end in #222 and at the container end in #276. These are **not synchronized terminal states**, a capacity measurement or proof of SOC error.

This timestamp qualification matters throughout the case: in #198, the last SOC records for R3, R4, R5, R7 and R9 precede the container end by more than six minutes. [The member table](member_summary.csv) retains every member, both SOC timestamps and the omitted-time accounting.

## Interpretation: why an asset owner might care

A difference becomes operationally useful only if it changes a specified judgment. The present case motivates three tests; it does not answer them.

| Asset question | What retaining the path makes possible | What is still needed |
|---|---|---|
| Can this request be sustained for the required period? | Inspect when current changes, how voltage responds, and which members differ during the period. | Requested power/current, actual limits and modes, aligned measurements and an explicit delivery criterion. |
| Why did the realized operation change or end? | Locate the change in the retained channels instead of attributing it to the endpoint summary. | Requests, limit reasons, alarms, PCS/BMS actions and termination records. |
| Does an earlier passage help interpret the next one? | Preserve member identity and the order of responses for a prospective comparison. | Comparable conditions, the missing intervening history, and evaluation using only information available at each decision time. |

For example, an owner might ask whether member histories help distinguish two requests that look alike in a summary but later differ in delivery. A valid test would compare a summary-only observer with one retaining the specified member history, against a predeclared observable outcome, without using future records. It must also check whether the existing BMS already retains equivalent information.

**The original BMS strategy is not available here.** These plots cannot show that it ignored the differences. Nor do they establish extra revenue, longer life, a safe limit, a fault, or a reason to change dispatch. Weihai is not a second-life validation dataset; that application would need its own evidence.

## How to read the evidence

- **Time:** each panel starts at its record-container boundary, not a verified physical onset. Gaps between containers are not observed rest. Native timestamps are retained; a timezone has not been established.
- **Current and voltage:** plotted at their own retained sample times, without smoothing, thinning or interpolation. Positive current denotes discharge; negative current denotes charge.
- **SOC:** the dataset’s reported estimate, not independent physical ground truth. Its first record need not coincide with the first current record; its last record need not reach the period end.
- **Ah:** each rack starts from its own first retained current point. Only adjacent current intervals no longer than 5 seconds are integrated, assuming linear change and separating charge from discharge at zero crossings. Larger gaps and unsupported prefixes/suffixes remain uncounted.
- **Support:** the partial Ah curve does not account for current in omitted intervals. Time coverage is not charge completeness. The downloadable records preserve the supported segments.
- **Comparison:** neither response waveform, temperature, actual request, control action nor prior history was matched. The three proximity rules are a way to choose examples, not proof of equal physical conditions.

## Check the case

[Member summaries and timestamp support](member_summary.csv) · [Plotted records and integral support](records.json.gz) · [Source, method and reuse notes](SOURCE.md)

The data originate from [Qu, Jingbo (2026), Dataset for: A Multi-Agent AI Framework for Explainable Battery System Maintenance](https://doi.org/10.5281/zenodo.20094787). These figures and examples are an LFR re-presentation, not an endorsement by the dataset author.

If a source value, timestamp or formation rule changes, the affected statement must be checked again. If the use changes from describing records to selecting actions, this case is not a substitute for that further test.

This case establishes only the stated differences among four retained record paths. No continuity pattern or causal difference is established.

### General observatory boundary, not an additional result of this case

Observation/measurement layer only. Not SOH, not RUL, not safety, not fault, not remaining-life, not warranty/underwriting/investment advice. It shows response-organisation continuity patterns visible in the named channels. Material-mechanism interpretation requires independent physical evidence (EIS / teardown / reference electrode / imaging).