Executive Summary

Physics AI system architecture at a glance

Table of Contents

Chapter 1: The Problem — Out of Sight, Out of Mind

Overhead distribution failures are visually obvious: a downed wire, a blown fuse, a tree on the line. Underground failures are invisible. Crews can spend hours with thumpers and TDR reflectometers narrowing down a fault. As a result, underground outage restoration is typically 3–4× more expensive than overhead and lasts longer. The trade-off is fewer outages, but each one costs more.

1.1 Scale of the trend

1.2 Why the analytics gap is widening

The grid is going underground faster than the data layer is keeping up. Utilities have SCADA at the substation and AMI at the meter — but everything in between is dark unless physically inspected. With aggressive undergrounding programs converting 5,000–10,000 miles per year nationally, the volume of "dark" infrastructure compounds.

1.3 Overhead vs underground failure profile

Chapter 2: Asset Universe — What's Actually Underground

An underground distribution feeder is not a single cable — it's a system of interconnected components, each with distinct failure modes and inspection regimes. A digital twin must model all six classes:

Chapter 3: Data Source Map

Every field rendered in the live demo maps to a real production data source. The table below is the master integration spec for a utility deployment. Status badges: Off-The-Shelf = utility almost certainly has it today; Sensor Deploy = requires a hardware install for full coverage (partial coverage from existing pilots is usually available).

3.1 Master integration table

3.2 Off-the-shelf data — the Day-1 stack

The first nine rows above are the foundation. Most utilities export GIS via Esri ArcFM (or modern Utility Network), keep OMS event logs in a SQL warehouse, and have at least 12–24 months of SCADA-historian data and AMI meter status. Earthflow ingests these as CSV / GeoJSON / SQL connectors and produces an AHI score for every cable segment within 2–3 weeks of data delivery.

3.3 Sensor-driven data — the Pilot stack

The four amber rows require hardware. The recommended pilot footprint:

• Sentient UM3+ underground line sensors — deploy 5–10 on the worst feeders the model flagged. Each unit covers up to 12 cable phases.
• CNIGuard / equivalent manhole IoT — 2–3 nodes in highest-risk vaults (downtown, near transformers, history of smoke).
• Online PD monitors — 1–2 systems on critical feeders or recently-replaced cable for QA.
• DTS fiber (optional, large utilities only) — co-located with cable in newer XLPE conduits.

3.4 The Environmental Intelligence Layer

A cable does not fail because of itself — it fails because of where it is buried. Everything above this point in the chapter is electrical data: what the utility already holds about the asset. This section covers the other half. The Orbyfy Environmental Intelligence Layer is built and maintained by Orbyfy, national and pre-built at 100% spatial coverage with 1,180+ data points at every lat/long, and resolved to the individual cable segment. The utility sources none of it; the only input required to unlock it is accurate route geometry. Two layers do the heaviest lifting for underground cable, and both are in production today:

• Soil corrosivity: drives lead-sheath degradation rate and grounding resistance. Our production pipeline ingests soil-property polygons — chloride and sulfate concentration, resistivity, pH and water-table depth — and joins them to cable segments by spatial intersection.
• Flood risk: 1%-annual-chance flood zones accelerate water-tree degradation in XLPE and lead-sheath corrosion in PILC. Cables in AE zones receive a +20% risk multiplier in the AHI weighting.

3.5 Which system serves which objective

The two objectives — Condition (how degraded, how likely to fail) and Capacity (how much load, how much headroom) — map cleanly onto the utility's own source systems, which makes it possible to say before a pilot starts exactly which extract powers which answer. C denotes Capacity; H denotes Condition.

GIS anchors both objectives — without route geometry neither question resolves per segment. Loading history (SCADA / AMI) is the bridge: it sets today's Capacity and, accumulated over years, drives Condition.

Chapter 4: Sensor Hardware Reference

A pragmatic survey of the devices a utility would actually purchase. The platform is sensor-agnostic — we ingest from any of these.

Chapter 5: Key Performance Indicators (KPIs)

5.1 Reliability metrics

5.2 Condition indicators

• Age vs Expected Life — age / design_life (40 yr XLPE, 50 yr PILC). Coarse but always available.
• Tan Delta (dissipation factor) — reveals dielectric losses; high = deteriorated insulation.
• Partial Discharge severity — pC magnitude + trend; the strongest leading indicator of failure within months.
• Load Factor — mean(load) / rated; sustained high load accelerates aging.
• Failure history — per-segment fault count in last 10 yr.
• Environmental risk — soil corrosivity tier × flood-zone multiplier.

5.3 Asset Health Index (AHI)

The composite KPI — details in Chapter 6.

5.4 Remaining Useful Life (RUL)

Survival-analysis estimate of years remaining. Anchored to AHI, modulated by trend in PD / load / thermal data when available. Output as years with confidence bands.

5.5 Capacity Index

The Capacity counterpart to the AHI. Where AHI and RUL answer how degraded is this asset and how long has it got, the Capacity Index answers how much more can it carry. It expresses, per segment, how much of the cable's dynamic thermal rating the measured load actually consumes — and therefore how much headroom remains for load growth, DER interconnection or a switching transfer. Details in Chapter 6.

Chapter 6: Physics AI Methodology & Model Portfolio

Earthflow's Physics AI engine is the analytical core of the underground twin. It is not one model — it is a coordinated portfolio of eleven production models, organized by what they serve — Condition (models 1, 2, 3, 5, 6, 7), Capacity (models 10, 11) and Operations & Planning (models 4, 8, 9) — that together turn raw utility data into prioritized, actionable risk intelligence. Each model is grounded in either a physics-based equation (corrosion kinetics, thermal aging, hydrology) or in an empirically-validated statistical method (survival analysis, time-of-arrival fault localization), then fused via the AHI composite. The result: explainable, defensible risk scores for every segment in the network.

6.1 The eleven production models

Models #1 (AHI), #2 (RUL), #5 (Soil), #6 (Flood), #8 (Optimizer), #9 (AMI Correlator), #10 (Dynamic Thermal Rating) and #11 (Loading & Headroom) all run from day-1 off-the-shelf data — no new sensors required. Models #3 (PD), #4 (Fault Triangulation), and #7 (Manhole Risk) require sensor data and come online in weeks 9–12 of the pilot. You get eight of eleven models running in your first week — and both indices, Health and Capacity, from day one.

6.2 The AHI Composite (Model #1) in detail

The AHI is the most important single model in the portfolio because it surfaces the highest-impact, easiest-to-act-on signal: a 1 (excellent) to 5 (end-of-life) composite score for every cable segment. It blends static attributes (age, material) with dynamic indicators (PD, load, fault history, environment) using a transparent, weighted-sum formula.

6.3 Sub-score scales

6.4 Worked example — Physics AI in action

6.5 The Capacity Index in detail

Condition tells a planner which segment is closest to failing. Capacity tells them which segment cannot absorb the next new load — a data centre interconnect, a fleet-charging depot, a rooftop-solar cluster reversing flow, or a switching transfer during a storm. Both questions arrive at the same planner's desk, and until now only one of them had an answer grounded in the asset's real physical environment.

The Capacity Index is computed in two stages. Model #10 (Dynamic Thermal Rating) establishes what the segment can actually carry. Nameplate ampacity assumes a reference installation; real segments deviate from it substantially. Neher–McGrath is evaluated per segment using the site's own soil thermal resistivity, burial depth, and whether the run is direct-buried or in duct — the same environmental layer that drives the corrosion and flood models. Model #11 (Loading & Headroom) then measures that rating against observed loading from SCADA feeder current and AMI interval data, producing percent-used and percent-headroom per segment.

Like every other output in the portfolio, the Capacity Index carries a value, a confidence interval, and a provenance record naming which inputs were measured and which were imputed — so a rating that rests on a modelled soil thermal resistivity rather than a utility survey says so explicitly.

Chapter 7: Rapid Deployment Roadmap (7-Day Quick-Start · 90-Day Pilot)

Earthflow is engineered for frictionless onboarding. There is no on-prem install, no infrastructure procurement, no 12-month enterprise rollout. A utility can be looking at Physics AI risk scores for their actual network in one week, and have a full sensor-validated pilot in three months.

7.1 Quick-Start: 7 days from data to dashboard

7.2 Full pilot: 90 days from quick-start to closed loop

After the 7-day quick-start, the full pilot extends to validate predictions with sensors and refine the models. Each step below assumes one utility data engineer + one Earthflow integration engineer in parallel.

Chapter 8: ROI Model

8.1 Value drivers

8.2 Worked 5-year ROI (mid-size utility, 50 feeders)

Larger utilities see proportionally larger absolute savings. The Dominion / FPL / PG&E case studies cited in industry reports show 20%+ SAIDI improvements and capex deferrals north of $1.76 per $1 spent on the analytics layer.

Chapter 9: Competitive Landscape

Four major vendors offer overlapping but differentiated solutions. Earthflow's positioning: the open, geospatial-first platform that complements rather than replaces existing OT/EAM stacks.

Chapter 10: Earthflow Capabilities Mapped to This Use Case

Every chapter of the live demo corresponds to one or more data sources and one or more user personas. The mapping below is the customer-facing rosetta stone for stakeholder briefings.

Chapter 11: FAQ for Utilities

11.1 What data is already in place before we ever talk to the utility?

Earthflow Underground starts with a data foundation we have already gathered, cleaned, and kept current — the Orbyfy Environmental Intelligence Layer. It is national and pre-built at 100% spatial coverage, carries 1,180+ data points at every lat/long, and is resolved to the individual cable segment. Nothing on this list requires anything from the utility — it is all in place on the day the conversation begins, and the only input needed to unlock it is accurate route geometry. Each dataset feeds one or more of the eleven Physics AI models, listed here alongside the answer it produces. Section 3.4 sets out the layer in technical detail; this is the plain-English version.

Every input carries provenance metadata. For any score we report, we can show the utility exactly which dataset contributed and where we used an industry default instead of the utility's own history.

11.2 What does the utility need to provide — and how do we handle it?

We work with whatever the utility can provide, in whatever shape it is in. Underground distribution data is rarely clean — that is a structural feature of the asset class, not a utility-specific failing — and we never ask a utility to scrub its records before sending them to us. The table below lists the inputs that compound value fastest, the ones that sharpen results when they happen to exist, and the ones we explicitly do not want the utility spending effort on.

Bottom line: six of the eleven Physics AI models go live from the cable map and outage log alone — even when those records carry the typical underground-data gaps. The daily-load readings and smart-meter data light up two more — the Capacity pair, which convert measured loading into a dynamic thermal rating and a headroom figure — leaving three that genuinely require field sensors.

11.2b Delivery, security and governance

Underground asset data is sensitive, and a utility security team will reasonably ask how it moves before asking what it produces. The arrangement is deliberately narrow.

• No operational-technology connectivity. Transfers are one-way, read-only extracts — a historian or GIS export, or an SFTP drop — taken from the utility's IT or DMZ side. There is never a live link into SCADA, DMS or protective relays, and nothing is written back to any control system.
• Agreement before data. NDA and data-sharing agreement are executed before any extract moves.
• Encrypted in transit and at rest, over SFTP or the utility's existing secure-transfer mechanism.
• Outputs handled as Critical Energy Infrastructure Information (CEII), with access control and audit logging.
• Controls aligned to NIST CSF and IEC 62443. Distribution data is treated with CIP-grade rigour regardless of whether it falls within formal CIP scope.
• Meter data de-identified and aggregated to transformer or segment before it reaches any model — the load shape is what the physics needs, not the customer.
• Coordinate reference stated explicitly (CRS / EPSG) with the GIS extract; state-plane or WGS84 are equally workable.

The practical consequence is that a pilot is a data-transfer exercise, not a systems-integration project. Nothing is installed on utility infrastructure, no agent runs inside the utility network, and the utility's systems of record remain the systems of record throughout.

How we handle the data itself

Underground distribution data is heterogeneous, partial, and often decades stale. Our architecture assumes that on day one, not as an exception to be managed:

• Schema Auto-Mapper. Detects field-name patterns across the common cable-mapping, outage-log, and historian schemas — Esri ArcFM, Schneider Conduit Manager, generic geodatabases, and in-house variants. Produces a candidate mapping in hours, not months. We target 80% schema coverage in the first 48 hours; an analyst-led validation pass closes the remaining 20%.
• No common data model required. We do not ask utilities to conform to one canonical schema. Earthflow runs on a semantic knowledge graph — heterogeneous source systems plug in through the Auto-Mapper and are reconciled in graph space, not by force-fitting fields into ours.
• Built for bad data. When utility-specific records are missing or stale, we impute from international failure curves (IEEE and CIGRÉ) and the data foundation in 11.1, and surface the gap as an explicit confidence interval — not hidden behind a precise-looking number.
• Provenance per output. Every score traces back to the inputs that produced it — which dataset contributed, where industry defaults filled a gap, and how stale each input was. Engineers, regulators, and rate-case reviewers can audit the basis of any specific recommendation.
• We do the unstructured extraction. Partial-discharge test reports, manhole inspection logs, and other artifacts stored as PDFs, scans, or paper attachments are parsed by us. Utilities never need to digitize first.

Seven days from data drop to a live risk twin. That is the engineering choice that makes this possible.

11.3 Do you need historical data to identify alarms or similar events?

To start, no. Six of the eleven Physics AI models run on day one using the data foundation above plus the utility's cable map, and the two Capacity models come online as soon as any feeder-loading history exists. Where the utility's own failure history is missing, we fall back to internationally cited industry failure curves (IEEE and CIGRÉ standards). The utility receives a ranked, citeable list of its worst-5%-of-segments within seven days of the data drop — even with zero labeled failures on file.

For a defensible back-test and tighter accuracy, yes. Our primary success measurement — getting at least 70 of every 100 “worst-cable” predictions correct against the utility's own historical fault locations — only becomes possible once five years of outage history is in hand. With that history, the Remaining Useful Life model calibrates to the utility's own vintages and failure patterns. Without it, results are still produced; they simply carry wider confidence intervals, surfaced explicitly in every output rather than hidden behind a precise-looking number.

In short: historical data is what proves the worst-5% list is right. It is not what makes the list possible.

Chapter 12: Glossary & References

12.1 Glossary

12.2 References

• U.S. Department of Energy — Undergrounding Transmission and Distribution Lines Resilience Investment Guide (2024)
• T&D World — Distribution Goal: 50% Underground by 2040
• FPL Storm Secure Underground Program (SSUP) — case study and reliability outcomes
• Sentient Energy — UM3+ Underground Line Monitor product reference and field results (20%+ CMI improvement, 65% patrol-time reduction)
• OMICRON — Online Partial Discharge Monitoring of Power Cables
• Energy Magazine — IoT Technology Prevents Exploding Manhole Covers (Con Edison + CNIGuard)
• Siemens Xcelerator — Cable Analytics (10× reduction claim, 10-yr capex deferral)
• Hitachi Energy — Lumada APM press release (2023)
• GE Vernova — Asset Performance Management Software
• Schneider Electric — EcoStruxure Asset Advisor & ArcFM Conduit Manager