📊 statistics

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

TRACE-C is an auditable, strictly-prior rank-calibrated detector for multi-stream operational telemetry that aggregates residuals from local, dependence, and temporal channels via Fisher aggregation to identify joint anomalies, as demonstrated by its ability to prioritize Storm Atiyah while revealing the distinct contributions of its components and clarifying the interpretive limits of its rank-based p-values.

Original authors: Matthew Faucher

Published 2026-08-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Matthew Faucher

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The electrical grid is a living, breathing system that never truly sleeps. It balances the constant hum of national demand against the fluctuating supply of wind and solar power, all while maintaining a precise rhythm known as frequency. When this system operates normally, its various signals—how much power is being used, how much is being generated, and how steady the frequency remains—move in a predictable, familiar pattern. However, the true challenge for engineers is not spotting a single number that looks strange, but recognizing when the entire pattern of these signals shifts in a way that is unusual, even if every individual number stays within its usual limits. It is like noticing that a familiar room feels different because the furniture has been rearranged, even though every piece of furniture is still the same size and shape. Detecting these subtle, collective shifts in real time is critical for preventing blackouts and managing the grid, yet it is a difficult task because the data is noisy, changes with the seasons, and is deeply interconnected.

A researcher has developed a new method called TRACE-C to solve this specific problem. Instead of trying to predict the future or guess what a "normal" day looks like, the system acts as a strict observer that only uses information available up to the moment it is making a judgment. It watches the grid's behavior in two-hour windows, comparing each new window against a history of similar windows from the past. The method is designed to be transparent and auditable, meaning every step of their decision-making process can be traced and verified. It does not rely on complex probability models that assume the data behaves in a perfectly random way, which is rarely true for real-world power grids. Instead, it uses a three-part check to see if a window of time is behaving strangely. First, it looks for sustained high or low values in any single stream of data. Second, it checks if the relationship between different streams has changed, such as when wind generation and demand usually move together but suddenly diverge. Third, it looks for sudden, sharp jumps in the data that break the expected flow from one moment to the next.

The researcher tested this system on real data from the Great Britain electricity grid, covering a period from early 2019 through the end of 2020. They split their work into two phases: a development phase where they tuned the system using data from 2019, and a frozen hold-out phase where they applied the exact same settings to 2020 without making any changes. The goal was to see if the system could identify genuine operational anomalies, such as those caused by major storms or unexpected power cuts, without being fooled by normal seasonal changes. The results were revealing and nuanced. When the system analyzed the 2019 data, it correctly identified a period of severe weather known as Storm Atiyah as the most unusual window of the year. However, a deeper look revealed that this detection was driven primarily by the first part of the system's check—the sustained high values in the data—rather than the part designed to detect complex relationships between different signals. In fact, if the researcher had relied only on the relationship-checking part of the system, Storm Atiyah would have been ranked much lower, around the 59th most unusual window.

The system also struggled to detect a specific, short-lived event that occurred on August 9, 2019, which involved a sudden drop in power and a disturbance in frequency. While other methods that focus on reconstructing what the data "should" look like ranked this event as the most important anomaly, the TRACE-C system ranked it much lower. This happened because the system's method of combining its three different checks tended to smooth out brief, sharp spikes in favor of longer, sustained periods of unusual behavior. The researcher found that the system's ability to detect short events was actually better when looking at just the third part of its check—the sharp jumps—rather than when all three parts were combined. This suggests that while the system is excellent at spotting broad, multi-day weather events that affect the whole grid, it may miss very brief, intense transients if they are not accompanied by a longer period of unusual behavior.

When the system was applied to the frozen 2020 data, it selected no alerts at all. This was not because the year was uneventful; 2020 contained major storms and the significant shift in demand caused by the pandemic lockdown. Instead, the lack of alerts was a sign that the system had become "saturated." Because the system keeps a growing history of every unusual window it has seen, once a very extreme event enters that history, it becomes much harder for any subsequent event to look even more unusual by comparison. The system did rank several windows from 2020 as highly unusual, including those corresponding to Storms Ellen and Alex, but none of them were extreme enough to trigger an alert under the strict rules the researcher set. This outcome highlights a fundamental trade-off in anomaly detection: being sensitive enough to catch every small problem often leads to too many false alarms, while being strict enough to avoid false alarms can cause the system to miss real events that are unusual but not record-breaking.

The study concludes that the TRACE-C system is a valuable tool for ranking and ordering potential anomalies, but it is not a magic bullet that can perfectly identify every problem. The researcher was careful to clarify that the scores the system produces are not probabilities of an event happening, but rather a way to sort windows of time from most to least unusual. They also emphasized that the system's ability to detect relationships between data streams is not based on a literal mathematical model of probability, but rather on a simplified algebraic form that works well in practice. By being honest about what the system can and cannot do, and by providing a complete record of their decisions, the researcher has created a tool that operators can trust to flag the most significant shifts in the grid, while understanding that some short-lived or complex events may require different methods to catch. The work serves as a reminder that in the complex world of operational monitoring, the best approach is often a combination of strict rules, transparent logic, and a clear understanding of the system's limitations.

Technical Summary: TRACE-C

Problem Statement

Operational monitoring of multivariate telemetry (e.g., electricity grid demand, generation, and frequency) faces a fundamental challenge: individual data streams may remain within historically familiar ranges while their joint configuration becomes anomalous. Conversely, seasonal or regime drifts can inflate conventional anomaly scores without indicating a discrete operational event.

Existing methods struggle with three specific constraints:

Dependence and Non-stationarity: Telemetry is serially dependent and non-stationary, violating the exchangeability assumptions required for sharp rank arguments.
Temporal Resolution: Coarse summaries can erase short physical transients, even if they make longer multi-stream episodes rankable.
Selection Policy: An operational detector is not just a scoring function but includes a selection policy (e.g., multiple testing rules, fallback mechanisms, attention budgets) that must be reported separately from the score.

Methodology: TRACE-C

TRACE-C (Rank-Calibrated Relational Anomaly Detection) is an auditable, strictly-prior detector designed for aligned multi-stream telemetry. It operates on non-overlapping windows (e.g., 2-hour blocks) and ensures all reference data is strictly prior to the scored window.

1. Regime-Conditioned Robust Residuals

The method partitions data by "regime" (settlement period crossed with weekday/weekend status). For each stream and regime, it computes robust $z$ -scores using the median and Median Absolute Deviation (MAD) of the $K=40$ most recent same-regime values.

Magnitude Preservation: Unlike rank-PIT (Probability Integral Transform) methods which cap magnitude, TRACE-C retains the magnitude of deviations up to a clipped limit ( $\pm 10$ ), avoiding the loss of information regarding the severity of physical excursions.

2. Three-Channel Scoring

The detector aggregates three distinct channels, each producing a score for a window $I_w$ :

Local Channel ( $L_w$ ): The maximum absolute normalized sum of residuals within a single stream. This captures sustained magnitude in at least one stream without diluting it via summation across streams.
Relational Channel ( $G_w$ ): A Gaussian copula-form negative log-density contrast. It uses a pre-fitted correlation matrix $R$ on the robust residuals. Crucially, no probability-integral or normal-score transform is applied to the inputs; the method applies the algebraic form of a copula contrast to robust- $z$ residuals, making it a "copula-form" score rather than a literal copula density.
Temporal Channel ( $T_w$ ): The worst (maximum) standardized innovation within the window, derived from per-stream AR(1) models. This preserves sharp transitions that might be diluted by averaging over a window.

3. Rank Calibration and Aggregation

Trailing Ranks: For each channel, the score is converted to a $p$ -value ( $p_{c,w}$ ) by ranking it against a trailing set of 240 strictly-prior scores.
Fisher Aggregation: The three channel $p$ -values are aggregated using Fisher's method ( $S_w = -2 \sum \log p_{c,w}$ ). The paper explicitly notes this is an aggregation score, not a chi-square test, as the channel ranks are dependent.
Outer Rank: The aggregate score $S_w$ is ranked against a growing set of all strictly-prior aggregate scores to produce the final outer rank $p_w$ .

4. Selection Policy

The selection process follows a strict hierarchy:

Nominal BH: An attempt is made to select windows using the Benjamini–Hochberg (BH) procedure at $q=0.05$ . The paper acknowledges that independence/PRDS conditions are not established, so this is a "nominal" attempt, not a guaranteed FDR control.
Record Fallback: If BH selects zero windows, the system falls back to an upper-record rule ( $p_w \leq 1/(|S_w|+1)$ ), selecting only windows that strictly exceed all previous aggregate scores.
Attention Budget: Selected windows are grouped into fixed 48-row blocks, retaining at most two alerts per block.

Key Contributions

Auditable Three-Channel Detector: A specification that preserves residual magnitude while ensuring all references are strictly prior.
Clarification of "Copula" Usage: Distinguishes the implemented Gaussian copula-form dependence contrast (applied to robust- $z$ residuals without marginal transforms) from literal copula calibration.
Transparent Selection Path: Reports the full selection logic, including the nominal BH attempt, the disclosed record-rule fallback, and the fixed-block budget, while explicitly stating the assumptions that prevent nominal FDR claims.
Frozen Hold-Out and Ablation: Provides a frozen 2020 evaluation and a 2019 channel ablation study. The ablation reveals that the top-ranked event (Storm Atiyah) is driven by the local channel, not the copula channel, and that Fisher fusion can bury short temporal events relative to single-channel rankings.
Evidence-Bound Package: The evaluation is generated from committed machine-readable reports with checksummed inputs, ensuring reproducibility without manual transcription.

Results

The evaluation uses Great Britain grid telemetry (2019 development, 2020 hold-out).

2019 Development:
- Storm Atiyah: Ranked 1st (p=0.000350). However, ablation shows this rank is local-led (copula-only ranks it 59th).
- 9 August Frequency Event: A short power cut is ranked 143rd by the fused detector but 40th by the temporal channel alone. Reconstruction baselines rank this event 1st. This indicates the fused detector buries short transients in favor of sustained multi-stream extremes.
- Selection: The record rule selected 2 alerts (Atiyah and an unlabelled window); BH selected none.
2020 Frozen Hold-Out:
- Zero Alerts: No windows passed the record rule. This is attributed to "record-rule saturation" (once a high score enters the history, the threshold becomes unreachable) rather than an uneventful year.
- Post-Hoc Interpretations: The highest-ranked window (20 August) was later interpreted as Storm Ellen (p=0.000504), and the 5th ranked window (3 October) as Storm Alex. Neither triggered an alert.
- Lockdown: The best window during the March 2020 lockdown was ranked 45th, not the onset window.
Baseline Comparison:
- Reconstruction methods (Autoencoder, PCA) outperformed TRACE-C on the short frequency event (Rank 1 vs. 143).
- TRACE-C outperformed baselines on Storm Atiyah (Rank 1 vs. 34 for Autoencoder).
- Isolation Forest performed best on the lockdown interval.
- No single detector dominated across all event types, highlighting an interaction between score construction, interval length, and sensor set.

Significance and Claims

The paper positions TRACE-C not as a universal anomaly detector or a literal copula calibration procedure, but as a strictly-prior rank-calibrated detector.

Interpretation of P-values: The output $p$ -values are selection quantities (upper-tail conformal/rank values) used for ordering and selection. They are not event probabilities, posterior confidence, or guarantees of time-series coverage.
Limitations: The author emphasizes that the method does not establish FDR control due to unproven dependence conditions. The "copula" name is historical/algebraic, not a claim of fitted density. The zero-alert result in 2020 is a substantive finding of saturation, not evidence of normalcy.
Honesty Ledger: The paper's primary contribution is the transparency of its operating envelope: it admits where it succeeds (sustained multi-stream extremes like Atiyah), where it fails (short transients like the frequency event), and the specific mechanisms (local channel dominance, record saturation) that drive these outcomes.

The work concludes that the appropriate interpretation is an analysis of the trade-offs between score construction, aggregation, and sensor resolution, rather than a universal ranking of algorithms.