TRACE-C: Rank-Calibrated Relational Anomaly Detection for Multi-Stream Operational Telemetry

TRACE-C: Rank-Calibrated Relational Anomaly Detection for Multi-Stream Operational Telemetry Matthew Faucher Independent researcher

arXiv:2608.21251v1 [cs.LG] 21 Aug 2026

19 Aug 2026

Abstract

is extreme, but whether a window is unusual relative to the information that would actually have been available when it arrived. That question is difficult for three reasons. First, telemetry is serially dependent and nonstationary, whereas simple rank arguments are sharpest under exchangeability [19]. Second, coarse summaries may erase short physical transients even when they make longer multi-stream episodes easy to rank. Third, an operational detector includes a selection policy as well as a score. A nominal multiple-testing procedure, a fallback rule, and an attention budget make different claims and must be reported separately. TRACE-C retains its historical name, “temporal relational anomaly detection with copula calibration,” but this manuscript narrows that description to the shipped algorithm. The method performs strictly-prior online scoring with rolling robust-z marginals and rank calibration. Its relational channel has the algebraic form of a Gaussian copula log-density contrast [21], but it is evaluated on robust-z residuals without a probability-integral or normal-score transform. It is therefore not a literal copula density. We keep the TRACE-C name for continuity and use “rank-calibrated relational anomaly detection” as the scientifically accurate description. The output p-values have an equally limited interpretation. They are upper-tail conformal/rank values used to order and select windows. They are not the probability that an alert is genuine, not posterior confidence, and not a guarantee of time-series coverage. Exchangeability provides a useful benchmark for interpreting ranks, but it is not established for these data. This paper makes five contributions:

Operational telemetry can be jointly anomalous while every individual stream stays inside its familiar range. TRACE-C is an auditable strictly-prior rank-calibrated detector for aligned multi-stream telemetry: sameregime rolling median/MAD residuals feed three window channels—a maximum normalized local sum, a Gaussian copula-form dependence contrast on robust-z residuals, and a worst standardized AR(1) innovation—whose channel ranks are Fisher-aggregated and ranked against earlier aggregates. We evaluate six Great Britain grid streams with a January–April 2019 fit, July–December 2019 development evidence, and a 2020 hold-out frozen before inspection. TRACE-C ranks Storm Atiyah first among 2019 test windows, but a disclosed channel ablation attributes that rank to the local channel, not the copula-form channel: copula-only ranks Atiyah 59th. The short 9 August frequency event is ranked far lower by the fused detector (143) than by the temporal channel alone (40), and reconstruction baselines rank it first. In 2020 no window is selected, which is consistent with record-rule saturation rather than an uneventful year; the highest-ranked frozen window was later interpreted as Storm Ellen. Three interpretive limits carry throughout. The resulting p-values are selection quantities, not event probabilities. The copula-form channel is not a literal copula density: the method applies no probability-integral or normal-score transform. Empirical rank counts are diagnostics, not coverage or false-discovery proofs. Every table and figure in this paper is generated from committed machine-readable reports.

1

1. It specifies an auditable three-channel detector that preserves residual magnitude while making every reference strictly prior to the scored window.

Introduction

Operational monitoring is intrinsically multivariate. Demand, embedded generation, pumping, and frequency can each remain within a historically familiar range while their joint configuration changes. Conversely, seasonal or regime drift can make many conventional anomaly scores large without a discrete operational event. The practical question is therefore not only whether a value

2. It distinguishes the Gaussian copula-form dependence contrast actually implemented from literal copula calibration, and distinguishes the Fisher aggregation score from a chi-square test. 3. It reports the full selection path—ordinary BH attempted first, a disclosed record-rule fallback, and a 1

fixed-block alert budget—together with the assump- [21]. TRACE-C borrows the resulting quadratic contrast tions that prevent a nominal FDR claim. but applies it to robust-z residuals. The absence of the marginal transform is material: “copula-form” describes 4. It reports a frozen 2020 run, a post-hoc non-identical the algebra, not a fitted copula density. baseline comparison, and a 2019 development-year Fisher’s statistic combines small p-values through a channel ablation showing that Atiyah’s rank-1 result sum of log terms [5]. Its familiar chi-square calibration is local-led rather than copula-only, and that Fisher requires conditions not asserted for TRACE-C’s depenfusion can bury the short frequency event relative to dent channel ranks. We use the construction only as an the temporal channel alone. aggregation score and calibrate that score again by its 5. It ships the evaluation as an evidence-bound package: online rank. Ordinary Benjamini–Hochberg (BH) controls FDR unthe TRACE-C hold-out was frozen before inspection, der independence and has known extensions under suitretained inputs are checksummed, and every table able positive dependence conditions [2]. Those indepenand figure in this manuscript is generated from comdence/PRDS conditions are not established here. General mitted machine-readable reports rather than trandependence motivates more conservative procedures such scribed by hand. as Benjamini–Yekutieli [3], but TRACE-C does not impleThe results precede the limitations and conclusion so ment that correction. Consequently, “BH” below means that the honesty ledger can be read against concrete a nominal selection attempt, not demonstrated FDR conevidence. TRACE-C ranks Storm Atiyah first in 2019 trol. development, but the ablation does not support a jointsurprise account of that window. Reconstruction methods isolate the short 2019 frequency event better. The frozen 2.3 Robust scaling, records, and compara2020 run produces no selected alerts. These outcomes tor families define the operating envelope supported by the present Median and MAD scaling provides resistance to extreme evidence. observations [17]. The fallback rule selects a new upper record; under an exchangeable continuous sequence its 2 Related work expected number of records follows the classical harmonic calculation [16]. Serial dependence, ties, and drift prevent 2.1 Rank calibration under dependence that expected count from being a guarantee in the present Conformal prediction obtains finite-sample rank guaran- application. The post-hoc comparison represents four common tees from symmetry or exchangeability arguments [19]. anomaly-score families: nonlinear reconstruction by an Operational telemetry generally has seasonality, autocorautoencoder [18, 23], PCA reconstruction residuals [6], relation, and changing regimes, so a correct arithmetic Isolation Forest [7], and the Spectral Residual method rank formula does not by itself supply those assumptions. [15]. These citations motivate the detector classes; the Methods for dependent data can instead use block permuexact fitted configurations and online rank envelope are tations or related constructions under explicit dependence described in Section 4. conditions [4]. TRACE-C does not implement the block method of Chernozhukov et al. [4]; the citation identifies a relevant alternative, not a theorem inherited by this 3 Method implementation. Conformal outlier testing connects rank p-values to multiple testing and can obtain FDR guarantees in a set- 3.1 Problem setup and online informating with independent reference and test samples and a tion set carefully characterized dependence structure [1]. That reference/test setting does not validate TRACE-C’s growing, Let xt,j denote aligned telemetry from stream j ∈ reused online reference. Here the ranks are transparent {1, . . . , d} at half-hour row t. TRACE-C partitions selection statistics and the observed-versus-benchmark the row sequence into non-overlapping windows Iw = {wW, . . . , wW + W − 1} with W = 4, corresponding to counts are empirical diagnostics. two hours on ordinary 48-period days. A score for Iw 2.2 Dependence scores and evidence ag- may use fitted parameters from the designated fit segment and observations with indices strictly smaller than the gregation row or window being scored. We call this strictly-prior Sklar’s theorem separates multivariate dependence from online scoring; it is an information-flow property, not a continuous marginals through probability-integral trans- causal-inference claim. forms [20]. Gaussian copula models then express depenThe desired output is a ranked list of anomalous windence through normal scores and a correlation matrix dows together with compact score contributions. No event 2

annotation enters the scoring function. The implemenThird, an AR(1) coefficient ϕj and innovation standard tation emits only the window index w, starting row t0 , deviation σe,j are fitted per stream on the same desigouter rank p, aggregate score S, channel contributions, nated fit segment. The temporal channel is the worst and per-stream contributions. It does not implement a standardized innovation within the window, separate severity field or a reason-code schema. |zt,j − ϕj zt−1,j | Tw = max . (4) j,t∈Iw σe,j

3.2

Regime-conditioned robust residuals

Taking a maximum rather than a mean preserves a sharp Each row is assigned a regime given by settlement period transition that would be diluted by the other rows of a crossed with weekday/weekend status. For every stream two-hour window. and regime, the method takes the K = 40 most recent same-regime values strictly before t, computes their me3.4 Trailing channel ranks and outer rank dian mt,j and median absolute deviation dt,j , and forms For each channel c ∈ {L, G, T }, let Rc,w be the trailing   xt,j − mt,j 240 eligible channel scores strictly before window w. Once zt,j = clip , −10, 10 . (1) 1.4826 dt,j the reference is full, the upper-tail channel value is MAD scaling follows the standard robust-scale construction [17]; a zero MAD is replaced in code by a small positive floor. This magnitude-preserving choice is deliberate. A sameregime rank-PIT with K reference values followed by an inverse-normal map cannot exceed Φ−1 ((K + 12 )/(K + 1)). At K = 40 the ceiling is approximately 2.2509, so a barely new record and a much deeper physical excursion receive the same transformed magnitude. Equation (1) retains magnitude up to the disclosed ±10 clipping limit; the later rank layers still govern selection.

3.3

pc,w =

1 + #{r ∈ Rc,w : r ≥ Cc,w } . |Rc,w | + 1

(5)

The three channel ranks are dependent because they summarize the same rows. TRACE-C constructs X Sw = −2 log pc,w , (6) c∈{L,G,T }

following Fisher’s aggregation form [5]. Equation (6) is only an aggregation score: no chi-square null distribution is claimed. The outer value ranks Sw against the growing sorted set Sw of all strictly-prior available aggregate scores. It is emitted after at least 40 prior S values:

Three window channels

The shipped detector contains exactly three channels. First, the local channel is the largest absolute normalized window sum,

pw =

1 + #{s ∈ Sw : s ≥ Sw } , |Sw | + 1

|Sw | ≥ 40.

(7)

All comparisons for selection use the exact arithmetic (2) value from Equation (7), not a displayed decimal. These are conformal/rank p-values for ordering and selection. Their finite-sample formula is exact, but the usual distriIt favors sustained magnitude within at least one stream butional interpretation still depends on exchangeability while avoiding a sum over streams that would conceal the [19]. responsible sensor. Second, let R be the normalized residual cross-product 3.5 Selection, fallback, and budget matrix estimated from complete robust-z rows in the January–April fit segment, with a ridge added if needed Across the scored windows in an evaluation segment, for a Cholesky factorization. The relational score averages the implementation first applies ordinary BH at nominal q = 0.05 [2]. Independence or PRDS for these reused  
1 online ranks is not established, so this step is only a 1 X 1 T −1 T Gw = zt R zt − zt zt + log |R| . (3) nominal selection attempt. If and only if BH returns zero W 2 2 t∈Iw windows, the code falls back to the upper-record rule 1 X Lw = max √ zt,j . j W t∈Iw

This is a Gaussian copula-form negative log-density contrast suggested by the usual Gaussian copula algebra [20, 21]. However, the inputs are the robust residuals in Equation (1); the implementation performs no PIT and no normal-score transform. Thus Gw is not a literal copula density, and its name is historical rather than a calibration claim.

pw ≤ pmin = w

1 , |Sw | + 1

(8)

which is equivalent to Sw strictly exceeding every earlier aggregate score. The fallback is commonly activated because rank granularity makes early BH thresholds unreachable. The report always names the rule used. 3

Under an exchangeable continuous score sequence, the record indicator has expectation 1/(|Sw | + 1) [16]. TRACE-C reports the sum of those expectations as an exchangeable-continuous benchmark, not as a guarantee for seasonal, autocorrelated telemetry. Finally, selected windows are grouped by fixed 48-row blocks and at most two are retained per block. This is an attention budget, not a true calendar-day or daylight-saving safe construction.

4

Data and evaluation protocol

4.1

Streams, aggregation, and provenance

hold-out. We use “frozen” to describe this development chronology, not to assert that the hold-out is a clean null year. In particular, 2020 contains major weather and demand-regime changes documented independently [9, 13, 22].

4.3

Annotations are joined only after ranking. The 9 August 2019 power cut and frequency disturbance is documented by Ofgem [14]; its evaluation interval overlaps 5 two-hour windows. Storm Atiyah spans 12 windows, while the two-day Ciara and Dennis intervals contain 24 windows each, using the Met Office storm chronology [9]. The first lockdown evaluation interval is 23 March–5 April 2020, or 168 windows, anchored to the 23 March announcement [22]. Because event intervals differ, their best ranks have unequal opportunity for a small value and should not be treated as like-for-like point-event metrics. Storm Ellen and Storm Alex were not pre-specified evaluation intervals. Their names were attached after ranking to the 20 August and 3 October windows using the external storm record [8, 9]. They are interpretations of unalerted windows, not detector inputs or discoveries.

The case study uses public Great Britain electricitysystem telemetry from the National Energy System Operator (NESO). Five demand-side columns are read at half-hour settlement resolution: national demand (ND), transmission-system demand (TSD), embedded wind generation, embedded solar generation, and pumped storage pumping [10]. A sixth column is formed from NESO system-frequency data by taking the maximum absolute deviation from 50 Hz within each settlement period [11]. This half-hour maximum retains an extreme deviation but not its sub-period timing or waveform. The retained 2019 and 2020 inputs are pinned by byte count and SHA-256 checksum. The report objects identify the generating source files and source-manifest hashes; the paper tables are generated from those committed reports. The data are used under the NESO Open Data Licence [12]. Required attribution is: “Supported by National Energy SO Open Data.” Rows are ordered by settlement date and period. The same-regime residual history distinguishes weekday from weekend and settlement period. The loader handles settlement-period irregularities defensively, while the later alert budget still groups fixed 48-row blocks; that distinction matters on daylight-saving transitions.

4.2

Event annotations and opportunities

4.4

Post-hoc baselines

Four baselines share the source rows, six-stream sensor set, W = 4 windows, fit cutoff, evaluation dates, score direction, and event intervals. They are a convolutional autoencoder inspired by standard reconstruction approaches [18, 23], PCA reconstruction with three components [6], a 200-tree Isolation Forest with fixed seed on window mean/standard-deviation features [7], and a trailing-context Spectral Residual score [15]. Their protocol is explicitly non-identical to TRACE-C. Each baseline uses global per-stream fit mean/standarddeviation scaling, then directly ranks its raw anomaly score against a growing strictly-prior history after a 40window post-fit warm-up. Selection is an unbudgeted record rule. TRACE-C instead uses same-regime robust residuals, three trailing 240-window channel ranks, Fisher aggregation, a growing outer rank, a BH-first selector, and a two-per-block budget. The baseline suite was constructed after TRACE-C’s 2020 results were known. Thus only TRACE-C was frozen before 2020 inspection, and Table 5 is a post-hoc comparison rather than a second hold-out experiment.

Segments and frozen protocol

January–April 2019 is designated as the fit segment for the dependence matrix and AR(1) parameters. It is assumed representative enough for this role; it is not established to be anomaly-free or “clean.” May–June supplies prior windows before the declared evaluation segment. July– December 2019 is the development segment, comprising 2,208 scored non-overlapping windows. Choices of W = 4, K = 40, the robust-z marginal, and inclusion of the frequency summary were informed by 2019 evidence and are not validation outcomes. The method, sensor set, and configuration were then fixed before the 2020 events were inspected. Rolling references continue across the year boundary, and all 4,392 scored windows in 2020 form the frozen TRACE-C

4.5

Reproducibility

The deterministic TypeScript evaluation rebuilds both TRACE-C reports from the pinned inputs. The baseline score artifact records its fixed random seed, trainer/source hash, input-series hash, runtime versions, and training history; the final evaluator verifies those fields before producing the comparison. Generated LATEX tables are 4

multi-stream extreme documented by the Met Office [9]. Table 1: Empirical rank-p diagnostics. Expected counts By contrast, the 9 August power-cut interval has a best are exchangeable-score references, not time-series coverage rank of 143 and p = 0.060420 across 5 opportunities, guarantees. despite the frequency-deviation stream. The source event Segment and threshold Observed/reference was real and operationally important [14]; importance 2019 development, p ≤ .05 120/110.4 and separability at the chosen aggregation are different 2019 development, p ≤ .01 23/22.1 properties. 2020 pre-COVID, p ≤ .05 50/45.0 2020 pre-COVID, p ≤ .01 2020 full year, p ≤ .05

3/9.0 205/219.6

5.3 derived from the JSON evidence rather than edited manually. These measures support computational reproduction of the reported arithmetic, while making no claim that different platforms will produce bitwise-identical externalmodel artifacts.

2019 channel ablation

Table 4 is a development-year diagnostic, not a holdout: the full detector’s Atiyah lead channel was inspected before the reruns. Isolating the copula-form channel does not recover Atiyah (rank 59). Dropping that channel leaves Atiyah at rank 2 without a record-rule alert, so the copula term changes whether the fallback fires, not whether the window is extreme. Local is the Atiyah engine 5 Results (local-only rank 3; dropping local removes alerts). The short frequency event is strongest on the temporal channel 5.1 Empirical rank diagnostics and selec- alone (rank 40 versus 143 fused). Fisher aggregation can therefore bury a brief transient that a single channel would tion rank more highly. These reruns do not support describing Table 1 compares the number of outer rank values below Atiyah as a copula-only discovery. selected thresholds with the arithmetic count αn. In 2019, the full report also gives 51 observed versus 44.2 at p ≤ .02. The displayed thresholds are close in some segments and 5.4 Frozen 2020 hold-out event ranks conservative in others: 2019 has 120/110.4 at .05 and 23/22.1 at .01; pre-COVID 2020 has 50/45.0 and 3/9.0. In the frozen 2020 run, Ciara’s best of 24 windows is The full 2020 count is 205/219.6 at .05 (and 39/43.9 at .01 rank 44 with p = 0.012058, and Dennis’s best of 24 is in the machine-readable report). These are descriptive rank 137 with p = 0.033217. These storm intervals follow empirical rank diagnostics. Autocorrelation, drift, reused the external chronology [9]. The lockdown annotation references, and known events rule out treating them as covers 168 opportunities from 23 March through 5 April proof of uniform p-values, coverage, or false-discovery [22]. TRACE-C’s best lockdown-window result is rank 45, control [1, 19]. Figure 1 shows the same diagnostics across p = 0.012559, at 06:00 on 28 March. It is not the onset window and must not be described as onset detection. all five detectors under the shared scoring envelope. The highest-ranked 2020 window, 20 August at 03:00, BH selected no windows in either reported TRACE-C was later interpreted as Storm Ellen; it has p = 0.000504 segment, so the disclosed record fallback ran. The 2019 but is unalerted. Rank 2, 20 January at 14:00, is unladevelopment segment contains two selected alerts versus belled. Rank 5, 3 October at 23:00, was later interpreted an exchangeable-continuous benchmark of 1.19; one is as Storm Alex using the external storm account [8]; it too the Atiyah window and the other is an unlabelled 21 July is unalerted. Table 3 lists the two post-ranking weather window. The frozen 2020 segment contains zero selected interpretations without implying that they are the top alerts versus a benchmark of 0.87. Neither comparison two windows or algorithmic discoveries. is a hypothesis-test guarantee [16]. The absence of later records despite several small rank values is evidence of record saturation: once a high aggregate score enters 5.5 Post-hoc baseline comparison the growing history, the rule becomes increasingly hard to trigger. Zero alerts is therefore a substantive frozen Table 5 reinforces the aggregation trade-off. The convoresult, not evidence that 2020 lacked unusual windows. lutional autoencoder and PCA reconstruction score the Figure 2 plots the full frozen segment with the post-hoc 2019 frequency event at rank 1, while Isolation Forest and Spectral Residual give rank 2. These detector families event annotations. target reconstruction, isolation, or spectral saliency rather than TRACE-C’s relational rank aggregate [6, 7, 15, 18]. 5.2 2019 annotated event ranks TRACE-C instead gives Atiyah rank 1; the autoencoder, The 2019 results are sharply mixed. Storm Atiyah is PCA, Isolation Forest, and Spectral Residual give Atiyah rank 1 of 2,208 with p = 0.000350 and passes the record ranks 34, 12, 3, and 26, respectively. rule. The full detector’s lead channel on that window For the 168-window lockdown interval, Isolation Forest is local, not copula-form, and the episode is a sustained ranks its best window first, PCA 22nd, the autoencoder 5

2019 test (Jul–Dec)

2020 pre-COVID (frozen segment) 2.07

2.0

2.0 1.45

1.5

1.82 1.78

1.5

1.31

1.09 1.11

1.09 1.04 1.0

1.11

1.11

1.0 0.64 0.68

0.64 0.59

0.60

0.59 0.59

0.5

0.5

conv-AE

IForest

PCA

SR

TRACE-C

0.44

conv-AE

0.33

IForest

PCA

SR

TRACE-C

dark bars: p ≤ .05; light bars: p ≤ .01; dashed line: observed = αn

Figure 1: Empirical rank diagnostics under the shared scoring envelope: observed counts of outer rank values at thresholds .05 and .01, expressed as a ratio to the arithmetic benchmark αn (dashed line). Values are read directly from the committed reports. Descriptive only: autocorrelation, drift, reused references, and known events preclude reading these as uniform p-value or false-discovery guarantees (Section 5). Dennis Ciara

COVID onset

S=27.4

S=0 Jan

Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Figure 2: Frozen 2020 segment: aggregate score S for each of the 4392 two-hour windows, scored with strictly-prior references under the configuration committed before 2020 data was examined. Shaded bands mark the post-hoc event annotations (Storms Ciara and Dennis; the COVID-19 lockdown onset). The record rule selected no 2020 window (Section 5); event windows rank in the upper tail without crossing the record threshold.

6

46th, and Spectral Residual 6th, compared with TRACEC at 45th. The best baseline dates are 5 April for the autoencoder, PCA, and Isolation Forest, and 28 March for Spectral Residual. These are interval search outcomes, not claims about 23 March onset. The pre-COVID .05 counts also vary widely: 27/45 for the autoencoder, 82/45 for PCA, 49/45 for Isolation Forest, and 93/45 for Spectral Residual. Because the baseline ranks use a global regime-naive reference and were produced after inspection of TRACE-C’s 2020 results, the table is neither a matched selection-budget leaderboard nor an independent validation.

Limitations and honesty ledger

The central limitation is calibration under dependence. The rank arithmetic is strictly prior, but grid telemetry is not shown to be exchangeable. The rolling 240-window channel references adapt to local score distributions; they do not turn the sequence into independent or identically distributed observations. The growing outer reference also reuses history across many tests. Consequently the empirical rank counts are descriptive, ordinary BH is nominal, and neither coverage nor FDR is established. Block conformal methods provide one relevant research direction [4], but no such block construction is implemented here. Likewise, the conditions studied for conformal outlier testing do not transfer automatically to this online reuse pattern [1]. The record fallback is transparent but weak on a long horizon. Its expected count is an exchangeable-continuous benchmark, not a time-series guarantee [16], and an early extreme can suppress later records. The zero-alert 2020 result alongside low-ranked unalerted windows empirically exhibits this saturation. A future selector should be specified before a new untouched period and evaluated with dependence and finite-resolution effects in view; this

Overall, no detector dominates. TRACE-C’s fused detector ranks Atiyah first, but the ablation shows that rank is not copula detection. Reconstruction is far better on the short frequency event, consistent with the temporal-only ablation ranking that event 40th while the fused detector ranks it 143rd. Isolation Forest supplies the smallest lockdown-interval rank. The appropriate interpretation is an interaction among score construction, interval length, sensor set, and temporal resolution, not a universal ranking of algorithms. 6

Table 2: TRACE-C event-window ranks. Labels are used only for post-hoc evaluation. Annotation

Segment

GB frequency event 2019 development Storm Atiyah 2019 development Storm Ciara 2020 frozen hold-out Storm Dennis 2020 frozen hold-out Lockdown transition (14 days; best 2020-03-28 06:00) 2020 frozen hold-out

Best rank

Best p Opportunity windows

143 0.060420 1 0.000350 44 0.012058 137 0.033217 45 0.012559

5 12 24 24 168

regime conditioning, two-stage ranks, aggregation, BH Table 3: Externally interpreted, top-ranked 2020 windows. attempt, and fixed-block budget. Only TRACE-C was Neither name was a detector annotation and no 2020 frozen before 2020 inspection, so apparent wins in either window passed the frozen record rule. direction are exploratory. Event-best ranks also favor External interpretation Rank Timestamp p longer intervals through more opportunities. Storm Ellen (post-ranking interpretation) 1 2020-08-20 03:00 0.000504 Storm Alex (post-ranking interpretation) 5 2020-10-03 23:00 0.001539 The attention limit is two selected windows per fixed 48-row block. It is not a calendar service: days with 46 or 50 settlement periods and daylight-saving transitions paper does not retroactively substitute such a rule. are not handled as true civil days. A deployment would Development history also limits inference. A histori- need a calendar-aware budget and explicit policy for late cal fixed-calibration episode produced 62 alerts; about or revised telemetry. 90% were judged seasonal-drift artifacts, corresponding Finally, the governance scope is deliberately narrow. to roughly 20-fold anti-conservatism. This is a narrative TRACE-C ranks process and sensor anomalies in operdisclosure only: there is no shipped artifact for that old ational telemetry for human review. It is not designed, run, and it is not a current comparator. The committed evaluated, or authorized for person-risk scoring, and the current comparison contains the demand-only variant, evidence in this study cannot support decisions about not a reconstruction of the historical fixed-calibration individuals. episode. Within the present algorithm, W ∈ {2, 4} and K ∈ {20, 40} were considered using 2019 development evidence, the change from rank-PIT to robust-z was 7 Conclusion event-informed, and the frequency stream was added after demand-only analysis. Freezing 2020 limits subsequent TRACE-C is best understood as a strictly-prior rankcalibrated detector, not a literal copula calibration proceadaptation but cannot erase those choices. The relational terminology needs similar restraint. dure. It combines magnitude-preserving robust residuals, Equation 3 resembles a Gaussian copula log-density con- local/dependence/temporal channels, a rank-aggregated trast, yet the robust-z inputs are not probability-integral score, and an explicit selection path. Its rank p-values transformed normal scores as in a literal copula construc- support ordering and selection; they are not estimated tion [20, 21]. The score may still be useful as a fitted event probabilities or posterior confidence. The evidence is mixed and operationally informative. dependence contrast, but copula likelihood theory does not directly calibrate it. The three channel ranks are also TRACE-C ranks Storm Atiyah first in 2019 development, dependent, so Fisher’s chi-square reference is deliberately but the channel ablation shows that this is a local-led window: copula-only ranks it 59th, and dropping the not used [5]. Sensor selection and aggregation define detectability. copula channel leaves it 2nd without a record alert. Simple The frequency stream keeps only the maximum absolute reconstruction, and the temporal channel alone, rank deviation within a half-hour, and all channels are eval- the short frequency event much better than the fused uated in non-overlapping two-hour windows. A short detector. The frozen 2020 run selects no alerts, while transient can lose its timing and be dominated by sus- its ranked list contains plausible weather-associated and tained background extremes, whereas a broad weather unlabelled extremes. The result exposes record saturation, episode can produce coherent contributions across several fusion trade-offs, and sensor/aggregation limits rather streams. The stream-contribution list discloses large win- than establishing nominal FDR or copula discovery. dow residuals but is descriptive rather than a mechanistic Future work should pre-register a dependence-aware explanation. A 2019 development-year channel ablation selector, evaluate it on a new untouched period, preserve is reported; it is peek-informed and is not a second hold- sub-period frequency information, test missing streams out. Missing-stream robustness, alternative aggregations, and alternative aggregations, and replace fixed 48-row synthetics, PR-AUC, and matched-budget comparisons attention blocks with calendar-aware budgeting. The curhave not been evaluated. rent contribution is the narrower one: an auditable algoThe baseline suite is post-hoc and protocol-non- rithm and evidence package whose assumptions, fallbacks, identical. Global scaling, direct growing raw-score ranks, and negative results are stated at the same resolution as and unbudgeted record selection differ from TRACE-C’s its successes. 7

Table 4: 2019 development-year channel ablation, not a hold-out. The full detector’s Atiyah lead channel was inspected before these reruns. Lower rank is better. Variant

Channels

Atiyah rank Blackout rank Alerts

full (L+G+T) local+copula+temporal drop copula local+temporal drop local copula+temporal drop temporal local+copula local only local copula only copula temporal only temporal

1 (alert) 2 3 5 3 59 10

143 69 119 338 241 323 40

2 1 0 1 0 0 0

Table 5: Post-hoc detector comparison. Lower event rank is better. The lockdown column searches a 168-window interval; parenthesized dates show each detector’s best window. Protocols are not identical and only TRACE-C was frozen before 2020 inspection. Detector

Blackout Atiyah Ciara Dennis Lockdown window Pre-COVID p ≤ .05

TRACE-C Convolutional autoencoder PCA reconstruction Isolation Forest Spectral Residual

143 1 1 2 2

1 34 12 3 26

44 1 12 10 12

Code and data availability The implementation, committed machine-readable reports, table generator, and reproduction instructions accompany this manuscript. Retained telemetry inputs are covered by checksums and report-generator provenance. Data are provided by the National Energy System Operator under the NESO Open Data Licence [12]. Required attribution: “Supported by National Energy SO Open Data.” Preprint: https://doi.org/10. 5281/zenodo.22012123. Code and reproduction instructions: https://github.com/mars-arch/trace-c. Evidence package: https://huggingface.co/datasets/ mars2titan/trace-c-neso-benchmark, DOI https:// doi.org/10.57967/hf/10063.

References [1] Stephen Bates, Emmanuel Candès, Lihua Lei, Yaniv Romano, and Matteo Sesia. Testing for outliers with conformal p-values. The Annals of Statistics, 51(1): 149–178, 2023. doi: 10.1214/22-AOS2266. URL https://doi.org/10.1214/22-AOS2266. [2] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300, 1995. doi: 10.1111/j.25176161.1995.tb02031.x. URL https://doi.org/10. 1111/j.2517-6161.1995.tb02031.x.

137 29 34 178 38

45 (2020-03-28) 50/45.0 46 (2020-04-05) 27/45.0 22 (2020-04-05) 82/45.0 1 (2020-04-05) 49/45.0 6 (2020-03-28) 93/45.0

[4] Victor Chernozhukov, Kaspar Wüthrich, and Yinchu Zhu. Exact and robust conformal inference methods for predictive machine learning with dependent data. In Proceedings of the 31st Conference on Learning Theory, volume 75 of Proceedings of Machine Learning Research, pages 732–749. PMLR, 2018. URL https://proceedings.mlr.press/v75/ chernozhukov18a.html. [5] Ronald A. Fisher. Statistical Methods for Research Workers. Oliver and Boyd, Edinburgh, 4 edition, 1932. URL https://archive.org/details/ statisticalmetho00fish_0. [6] J. Edward Jackson and Govind S. Mudholkar. Control procedures for residuals associated with principal component analysis. Technometrics, 21(3):341– 349, 1979. doi: 10.1080/00401706.1979.10489779. URL https://doi.org/10.1080/00401706.1979. 10489779. [7] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE International Conference on Data Mining, pages 413–422, 2008. doi: 10.1109/ICDM.2008.17. URL https://doi.org/10. 1109/ICDM.2008.17. [8] Met Office. Storm alex. Met Office, 2020. URL https://www.metoffice.gov.uk/weather/learnabout/weather/types-of-weather/storms/ storm-alex. [9] Met Office. Uk storm centre: 2019–2020 storm season. Met Office, 2020. URL https: //www.metoffice.gov.uk/weather/warningsand-advice/uk-storm-centre.

[3] Yoav Benjamini and Daniel Yekutieli. The control of the false discovery rate in multiple testing under dependency. The Annals of Statistics, 29(4):1165– 1188, 2001. doi: 10.1214/aos/1013699998. URL [10] National Energy System Operator. Historic https://doi.org/10.1214/aos/1013699998. demand data. NESO Open Data Portal, 8

2019. URL https://www.neso.energy/data- [20] A. Sklar. Fonctions de répartition à n dimenportal/historic-demand-data. sions et leurs marges. Publications de l’Institut de Statistique de l’Université de Paris, 8:229–231, [11] National Energy System Operator. System 1959. URL https://gallica.bnf.fr/ark:/12148/ frequency data. NESO Open Data Portal, bpt6k6116532p. 2019. URL https://www.neso.energy/data[21] Peter X.-K. Song. Multivariate dispersion modportal/system-frequency-data. els generated from gaussian copula. Scandinavian Journal of Statistics, 27(2):305–320, 2000. doi: [12] National Energy System Operator. Neso open data 10.1111/1467-9469.00192. URL https://doi.org/ licence. NESO Open Data Portal, 2024. URL 10.1111/1467-9469.00192. https://www.neso.energy/data-portal/opendata-licence. [22] UK Government. Prime minister’s statement on coronavirus, 23 march 2020. GOV.UK, 2020. URL [13] National Grid ESO. 2020 end of year review. National https://www.gov.uk/government/speeches/pmGrid ESO, 2021. URL https://www.neso.energy/ address-to-the-nation-on-coronavirus-23news/2020-end-year-review. march-2020. [14] Office of Gas and Electricity Markets. Energy [23] Ashwin Vijay. Time-series anomaly detecperspectives following the national electricity transtion using an autoencoder. Keras code exammission network power cut of 9 august 2019. Ofgem ple, 2020. URL https://keras.io/examples/ report, 2020. URL https://www.ofgem.gov.uk/ timeseries/timeseries_anomaly_detection/. publications/energy-perspectives-followingnational-electricity-transmission-networkpower-cut-9-august-2019. [15] Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang, et al. Time-series anomaly detection service at microsoft. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 3009–3017. ACM, 2019. doi: 10.1145/3292500.3330680. URL https://doi.org/ 10.1145/3292500.3330680. [16] Alfréd Rényi. Théorie des éléments saillants d’une suite d’observations. Publications of the Mathematical Institute of the Hungarian Academy of Sciences, 7, 1962. URL https://eudml.org/doc/80373. [17] Peter J. Rousseeuw and Christophe Croux. Alternatives to the median absolute deviation. Journal of the American Statistical Association, 88 (424):1273–1283, 1993. doi: 10.1080/01621459. 1993.10476408. URL https://doi.org/10.1080/ 01621459.1993.10476408. [18] Mayu Sakurada and Takehisa Yairi. Anomaly detection using autoencoders with nonlinear dimensionality reduction. In Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, pages 4–11. ACM, 2014. doi: 10.1145/2689746.2689747. URL https://doi.org/ 10.1145/2689746.2689747. [19] Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 9:371–421, 2008. URL https://jmlr.org/ papers/v9/shafer08a.html. 9

Related documents

• Real-time and adaptive anomaly detection algorithm for cyclostationary models#987087
• LLM-Enhanced Log Anomaly Detection: A Comprehensive Benchmark of Large Language Models for Automated System Diagnostics#13168
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability#1169667
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability#1156103
• Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection#238638