← Similar summaries. Different paths.

From State Summaries to Response Paths

A Traceable Observation Workflow for Field BESS Records: The Weihai Case

English web reading edition v0.1 · 25 September 2026 · Descriptive methods demonstration; not peer-reviewed.

This English methods account accompanies the interactive exhibit. It reuses existing records, methods and figures. Public preparation adds no measurements or experimental recomputation and does not extend the interpretation. Source and method identifies the dataset, transformations, included evidence and reproduction limits. The local draft is retained unchanged; this web edition replaces local links with the declared public evidence subset.

Abstract

Field battery records are often compressed into state estimates, interval statistics or health features. These representations are useful, but similarity in a summary does not make the member records and processes that produced it identical. Using existing records from nine racks in a Weihai battery energy storage system (BESS), this methods case presents an observation workflow that retains member identity, native timestamps, visible current passage and missing intervals. It starts with the full retained record, then examines specific relations, preserving a route from each summary back to the records that formed it.

Existing examples show two distinctions. Named-member voltage caches can change while a custom summary remains exactly the same. Current and voltage paths can differ when first reported state of charge (SOC), retained duration and supported ampere-hour subtotals are close. A further current-range comparison shows that similar current does not automatically establish the same stage of progression or a comparable state. A history-replay experiment with mixed results is also retained, separating “preserving distinctions” from “improving a judgement with those distinctions”.

The contribution offered here is an inspectable record-reading workflow with concrete examples, not a new battery-health criterion. It does not identify internal material mechanisms, prove that the original battery management system (BMS) overlooked member information, establish control benefits or reconstruct a complete physical history.

Keywords: field energy storage; asynchronous records; member identity; process paths; state summaries; observation boundaries.

1. The question: what does a summary represent?

If two charge or discharge records have similar starting SOC, duration and accumulated Ah, they are close on those coordinates. Further questions remain: does current rise early or late? How long is the higher current sustained? Where do member voltages begin to separate? Are there any retained records after an upward turn?

Those questions cannot be answered from the summaries alone. A summary is not necessarily erroneous data. It is a particular compression of a process. What needs qualification is the kind of similarity that the compressed representation supports.

The question is:

When field BESS records are compressed into state summaries, which member and process differences cease to be visible? How can those distinctions be retained without interpreting the records as physical facts they have not established?

“Method” here means an observation and evidence-organisation workflow. It is not a new score replacing SOC or SOH, nor a claim that drawing more curves makes an analysis superior to existing algorithms.

2. What the data can observe

2.1 Object and scope

The public Weihai dataset comes from an operating energy storage system. This document reuses rack current, terminal voltage and, in selected cases, original BMS SOC records already extracted and retained by the project. Members are distinguished by their recorded identities, R1–R9. The observation unit is each member's retained process within the same system. Racks are not treated as independent battery samples, and individual points in a long record are not treated as independent experiments.

The release is partitioned into charge/discharge operations. Here these are called “retained segments” or “records”, avoiding the implication that physical task onset, termination and the intervening history were fully observed. File month, record-container boundaries, each channel's timestamps and physical action boundaries are not equivalent. No unrecorded experience between containers is reconstructed. [S1][S2]

Only previously analysed material is reused. The interactive public replay contains four selected records, not the full dataset. Record-by-record reading of the full open range remains unfinished. The cases were selected after exploration; they are not a blind validation set and do not estimate the prevalence of a phenomenon across the dataset. [S9]

2.2 Channels and coordinates

Record or coordinate Use here What it does not automatically represent
Rack current I Recorded actual current; negative for charge, positive for discharge External requests, original BMS decisions or the complete unrecorded current history
Rack terminal voltage V A terminal-voltage projection retained at native timestamps Internal material state, member health or a synchronous whole-system state
Original BMS SOC A reported estimated coordinate, with source time and record age Independent ground truth, remaining energy or the validation target for a different observation method
Ah subtotal Integration over declared intervals between adjacent current records Capacity, all charge that passed or a common physical progression coordinate across members
Time since record start A navigation coordinate within the stored record Known physical onset or continuous operating duration
Gaps within or between records Lack of record support for that channel Zero current, rest, no change or physical recovery

The materials used here do not supply a complete chain of external requests, current limits, balancing, contactor actions and termination decisions. Temperature is not used in common matching or attribution in the cases below; this does not mean temperature is unimportant. The separate contributions of the battery, environment, control and measurement cannot be isolated here. [S1][S2][S3]

3. Method: retain a route from summary to process

Existing retained records and their sources
        ↓
Member identity, native timestamps, support and gaps
        ↓
Current–voltage paths across the full retained segment
        ↓
Declared comparison: what is close, different or unknown
        ↓
Member distances, ordering and locations of change
        ↓
A bounded description + explanations still open

This is a reading sequence, not a new data-generation chain. It does not imply that every example implements every step.

3.1 Keep each channel on its own clock

Every value used should retain at least record identity, rack, channel, source row index, source time and value. If a display cursor uses the latest preceding record, its age should also be retained.

A set of values at one display cursor is a collection of records selected by a rule. Its source times may differ. It cannot simply be called the common physical state of the battery at that instant. Connecting points, interpolating or using finer playback steps does not increase the temporal resolution of the original observations.

3.2 Show the full retained segment before inspecting a detail

First show the native paths of every member retained within the container, then mark local spans. A detail view must lead back to the full record so that its lead-in, subsequent records and missing intervals do not disappear from interpretation.

“Full” always means the full retained span, not a complete continuous physical experience. Use interpretable shared plotting scales across records. Do not stretch unequal durations into a common percentage and silently call it the same physical progress.

3.3 Declare the comparison conditions

“Similar records” must specify the coordinates on which similarity is defined. Direction, first records, actual current paths, elapsed duration, visible charge subtotals, the SOC coordinate and record support should be recorded separately as relevant to the question. Unobserved conditions remain unknown.

Matching a subset of conditions can support a descriptive comparison under that subset. It cannot establish identical complete initial states or causal attribution. Nor must every unknown be resolved before looking at a figure: an unknown limits only the conclusions that depend on it.

3.4 Put summaries beside their member records

Let C(t) be a member cache formed in archive-time order using only preceding records, and let S(t)=g(C(t)) be a declared summary. We inspect whether there is an instance with:

S(a) = S(b), while the named-member voltage caches C_V(a) ≠ C_V(b).

Such an instance shows that g is not a unique representation of the retained member records. It does not establish that a and b are synchronous physical states, or that the cached difference requires intervention.

Member relations should retain distance as well as ordinal position, together with the sources of both operands. For example, D_ij = V_i(t_i) − V_j(t_j) must retain t_i, t_j and the acceptable record-age rule. It must not silently become an unknown synchronous physical voltage difference. When order changes near a quantisation step, measurement and asynchronous recording remain possible contributors.

3.5 Account for charge passage only over supported intervals

The process cases below inherit an existing Ah rule: integrate only between consecutive retained current records separated by no more than 5 seconds, assuming linear current within each interval. Split charge-in and discharge-out when current crosses zero. Do not integrate across larger gaps or fill unobserved prefixes, suffixes or inter-record blanks. [S3]

Each rack's ledger begins at its own first retained current record. A subtotal can accumulate disjoint supported intervals, but the omitted intervals must remain identifiable. The 5-second parameter is an existing display scenario, not a guarantee of physical integration accuracy.

3.6 Separate visibility from practical benefit

Member identity makes some distinctions locatable. Full retained paths make the location of differences within a process visible. Neither automatically improves estimation, warning or control.

A claim of improvement for a particular judgement requires a separately declared target, baseline, usable past information and evaluation scope. Until then, “more information” cannot substitute for utility testing. The existing result that did not show universal improvement is retained as an example of this limit.

4. Observed: demonstrations in existing records

4.1 Identical custom summaries, different named-member records

An existing offline replay uses charge records #1/#349 and discharge records #5/#350. Native I/V/SOC records enter a cache in archive-time order; records with the same timestamp enter as one batch. The fixed summaries are median V, V range, median I and median SOC across all nine members. Incomplete channel coverage is not replaced with a partial-member summary. Equality uses unrounded values. [S4]

Each of these retained segments contains an instance in which the summary is unchanged while one or more named voltage-cache values change. Figure 1 shows the first qualifying instance in each segment under the existing rule, not the largest difference.

Figure 1: identical custom summaries with different named-member voltage caches

Figure 1. Open circles and filled points denote the cache before and after a record batch enters. Each value is recoverable through its named member, source index and timestamp. Display decimals do not determine equality; unrounded values do. This is a comparison of cache representations, not a synchronous physical-event plot or a comparison with the original BMS implementation.

The narrow observation is: this custom summary does not uniquely represent the named-member voltage records it compresses. The original BMS may use member channels and other protective paths. Its internal implementation was not tested here.

4.2 Similar summaries, different realised paths

An existing case selects discharge records #130/#198 after inspecting open records, with charge records #222/#276 retained as a second example pair. The selected coordinates are each rack's first SOC, retained duration and supported Ah. Temperature, control, full history and current waveform were not matched. [S5]

Figure 2: complete retained paths of discharge records 130 and 198

Figure 2. Each column retains I, V, reported SOC and supported charge passage for all nine members. Time is minutes since the stored record's start. Ah is not capacity; SOC is not independent ground truth. Overlapping native points do not imply continuous observation between them; the rows are not synchronous acquisitions.

In #130, current holds at a lower level before moving to a higher level. In #198, current reaches the higher level earlier and steps down near the end while voltage records move upwards. These differences are visible in the retained paths but are not uniquely determined by the selected summaries.

Actual current itself differs here. The case establishes that similar summaries are insufficient to confirm identical processes. It cannot establish that the physical battery changed under identical inputs.

4.3 A similar current range is not the same point in the process

An existing #195/#198 comparison fixed the navigation range 140 ≤ I < 150 A before computation, while retaining other bins and the full paths. Support within this range occurs at different locations in the two records, with different elapsed container times, supported charge subtotals and reported SOC positions. [S6]

That case also retains voltage-difference paths for every member pair. Different full paths do not imply that every relative order reverses. Preserved order does not imply unchanged spacing or process.

“Current is close again” can open a comparison; it does not complete the condition check. Joint changes in visible conditions do not identify the causal contribution of any one of them.

4.4 History did not universally improve record reconstruction

An existing member-history replay learns separately within each rack and retained segment. It estimates the current V record from information available beforehand, then reveals that record and updates. It compares a current-history baseline with representations that add recent V history, retaining history-order perturbations and alternative record-support scenarios. [S7]

The results do not support a conclusion that adding history is universally better. Improvement and worsening coexist. Some comparison directions reverse when record-support filtering changes. That filtering changes the trainable and scorable nodes together, so a reversal cannot be attributed solely to record-update speed.

This is a local, retrospective reconstruction of records at the same nodes. It is not an original-BMS performance comparison or a prospective deployment validation. It does not test the control value of the entire workflow presented here. Whether a particular history representation helps must be tested for a particular question.

5. Interpretation: what do these observations add?

5.1 A bounded, inspectable conclusion

In the selected Weihai records, retaining member identity, native record support and full paths lets us locate some record differences that selected summaries do not uniquely represent. Those differences have not been established as changes in internal state or as grounds for improved control.

Locating these differences depends on the original records still being retained. It is not recovery of lost information from summaries. Many-to-one compression is not a new discovery here. The demonstration concerns how this occurs in actual asynchronous BESS records and how to avoid silently changing the object being compared.

5.2 Relation to existing research and tools

The field-BESS preprint by Wong and colleagues extracts and normalises health features per cell group, explicitly accounting for operating conditions. Its Gaussian-process model adds an operating-condition term and a time term and declares an independence assumption between them. [R1] This document does not test that model or use its data. One resulting question is whether a member's response to operating conditions changes with historical position, and whether operating-condition/history interactions deserve a separate test. This remains a question, not an established finding here.

TWAICE product material already describes investigation using member-level historical curves, metrics and anomalous conditions. [R2] This document therefore does not claim that other teams look only at summaries or ignore processes. The proposed contribution is an inspectable connection among summaries, member paths, record support and bounded interpretation. Academic novelty still requires a fuller related-work comparison.

5.3 Possible practical relevance

The workflow can help formulate more specific checks: is the same spread formed by the same members? Does the same SOC representation accompany different recorded current histories? Does a local change occur within a well-supported interval or at a record boundary?

Such questions may change what should be observed or checked next. They do not themselves produce current-limit, shutdown, balancing, maintenance or replacement recommendations. Establishing those uses requires corresponding action records, an evaluation target and validation. Richer graphics do not supply that missing step.

6. Limitations and unfinished work

7. Closing

The practice demonstrated here keeps a state summary connected to the members and process it compresses. A summary has uses; the process has aspects the summary cannot replace. Putting them side by side lets us specify more carefully what is close, what differs, where a difference appears and which explanations remain indistinguishable.

The current deliverable is this observation workflow and its bounded examples. Whether it improves a specific judgement is a separate evaluation question. A demonstration of visibility must not be presented as an already established practical benefit.

References and supplied evidence

The related-work descriptions are inherited from the methods draft's source check on 25 September 2026. Public preparation does not test those approaches or add a literature review.

Observation/measurement layer only. Not SOH, not RUL, not safety, not fault, not remaining-life, not warranty/underwriting/investment advice. This account describes recorded response relations; it does not establish internal mechanisms, original-BMS omission or controller benefit.

General observatory boundary, not an additional finding

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).