Case Study

AI-Powered Grid Fault Intelligence: How a State Distribution Utility Cut AT&C Losses and Learned to See Its Own Network

Industry: Power & Energy | Location: Confidential

About the Project

The client is a state-owned electricity distribution utility serving a mixed urban-rural geography across two Indian states. Its network reaches over 4 million consumers through approximately 12,000 distribution transformers and nearly 85,000 km of LT/HT lines. SCADA coverage was limited, and operational data lived across multiple legacy platforms that rarely spoke to one another.

Kreate Technologies partnered with the utility to design and build the Grid Fault Intelligence and Predictive Distribution Health Platform, a purpose-built AI system trained on the utility's own operational reality, developed and refined over an 18-month engagement.

The Challenge

A Grid Flying Blind

A network of this size generates an extraordinary volume of signals every day: voltage readings, load patterns, meter data, fault reports, weather conditions. The utility was collecting a fair amount of it. What it lacked was a way to turn that data into foresight. As one senior director on the utility's operations team put it, the network was being run "with essentially spreadsheets and gut instinct." That wasn't a failure of the team; it was a natural outcome of running a rapidly growing, geographically dispersed distribution network on tools that were never designed to keep pace with its scale. The organisation needed a system that could see patterns its people couldn't, and act on them before damage was done. That gap showed up in four concrete ways:

Problem Statement

AT&C losses above 28%, well over the national benchmark of 15–18%, quietly draining revenue every quarter with no clear diagnostic path to bring them down.
Entirely reactive fault response. Field teams typically learned about failures only after consumers complained, with a mean time to detection of 4.2 hours, after the failure had already occurred.
Unquantifiable electricity theft. Non-technical losses were widely suspected but impossible to isolate at scale, since there was no automated way to separate genuine consumption anomalies from tampered connections.
Calendar-based maintenance. Field crews were dispatched on fixed schedules regardless of actual asset condition, which meant expensive interventions on healthy equipment and blind spots on the assets that were actually deteriorating.
None of this pointed to a lack of engineering capability. It pointed to a grid operating at a scale where manual monitoring and periodic maintenance schedules had simply run out of runway.

Our Solution

Kreate Technologies built the Grid Fault Intelligence and Predictive Distribution Health Platform as a modular AI system, four interconnected intelligence engines feeding into a single command interface used by control-room and field teams.

Predictive Fault Intelligence (PFI)

LSTM neural networks combined with gradient boosting analyse voltage fluctuation, harmonic distortion, load imbalance and thermal stress patterns that tend to precede physical failures. The engine produces a rolling 72-hour fault probability score for every monitored asset.

Why it mattered: For the first time, the utility could act on a failure before it happened, rather than reacting once a consumer had already lost power.

Non-Technical Loss (NTL) Detection AI

Isolation Forest models, combined with a clustering approach, build an expected consumption profile for each consumer segment. When actual usage diverges from that profile, the system assigns a prioritised suspicion score and generates action lists for field verification teams.

Why it mattered: Theft investigation shifted from broad, resource-intensive sweeps to targeted visits guided by evidence, making non-technical loss something the utility could finally measure and manage.

Transformer Health Scoring (THS)

The platform continuously tracks loading ratios, oil temperature trends and historical failure correlations, assigning each transformer a dynamic health score from 0 to 100.

Why it mattered: Maintenance shifted from a fixed calendar to actual asset condition, so crews spent their time on the transformers that needed attention rather than the ones that happened to be next on a schedule.

Renewable & Load Forecasting

Using weather APIs, solar generation data and consumer demand profiles, the platform generates 24–72-hour load and generation forecasts at the feeder level.

Why it mattered: As renewable penetration on the network grows, planners can configure the grid proactively rather than adjusting reactively after demand or generation shifts have already occurred.

Our Solution

Layer : Data Ingestion | Components : Smart meters, SCADA, AMR, weather APIs | Role : Real-time and batch collection

Layer : Edge Processing | Components : On-premise agents at substations | Role : Low-latency local alerts

Layer : AI Engine Layer | Components : PFI, NTL, THS, forecasting models | Role : Core intelligence and scoring

Layer : API Gateway | Components : RESTful APIs, role-based access | Role : Secure data orchestration

Layer : Command Interface | Components : Web and mobile dashboard | Role : Supervisor decision support

Layer : Integration Layer | Components : SAP, MDMS, billing systems | Role : Enterprise connectivity

Platform Architecture:

The platform was built as a cloud-native, API-first system with on-premise edge AI agents at major substations , keeping time-critical fault signals processed locally, while centralising model execution in a secure cloud environment for continuous learning across the network.

Under the Hood: How the AI Was Built

Data foundation first. Kreate Technologies spent the first three months of the engagement not writing models, but building a reliable, clean data pipeline, integrating 14 disparate sources, including legacy SCADA outputs, billing records, digitised fault registers, weather feeds and AMR reads. Every downstream model depended on this groundwork holding up.

Purpose-fit model selection.

Rather than applying a single general-purpose model across every use case, the team matched the model to the problem: LSTM networks for temporal fault recognition, XGBoost for transformer health scoring, an Isolation Forest and DBSCAN combination for non-technical loss anomaly detection, and a Prophet plus custom neural network ensemble for load forecasting, tuned to handle Indian seasonal and festive demand patterns. Layer Components Role Data Ingestion Smart meters, SCADA, AMR, weather APIs Real-time and batch collection Edge Processing On-premise agents at substations Low-latency local alerts AI Engine Layer PFI, NTL, THS, forecasting models Core intelligence and scoring API Gateway RESTful APIs, role-based access Secure data orchestration Command Interface Web and mobile dashboard Supervisor decision support Integration Layer SAP, MDMS, billing systems Enterprise connectivity

Explainability by design.

Every AI output carries a plain-language justification, not just a score, but a narrative covering overloading patterns, thermal stress probability, days since last inspection, and a recommended action. Field engineers engage with the system as a tool that supports their judgment, not an oracle that replaces it. That distinction turned out to matter more than any single model's accuracy.

The platform runs on a continuous feedback loop: deploy, capture field feedback through the mobile app, retrain monthly. That loop, not a one-time model swap, is what drove fault prediction accuracy from 74% at deployment to 91% over 18 months.

As the utility's chief engineer of grid operations described the shift: once the system began explaining why it was flagging something, rather than simply that it was flagging it, engineers stopped second-guessing it and started acting on it.

From Foundation to Full-Grid Intelligence

Foundation.

The first phase focused entirely on data, unifying 14 fragmented sources into a single reliable pipeline before any model went into production.

Pilot.

Initial models were deployed at 74% fault prediction accuracy, giving the utility its first working view of failures before they happened, and establishing the mobile feedback channel that would drive future improvement.

Expansion.

A dedicated Kreate AI trainer sat in the utility's daily operations meetings for the first six months, refining models, fixing pipeline edge cases and iterating the interface based on how supervisors and field engineers actually used it. Monthly retraining cycles steadily pushed prediction accuracy from 74% to 91%.

Scale.

Over 18 months, the platform moved from a pilot capability to the system of record for fault prediction, transformer health, non-technical loss detection, and load forecasting across the full network, with Kreate Technologies remaining embedded throughout the engagement rather than handing off a static tool.

Results

Business Outcomes

34% reduction in AT&C losses, down from 28.4% to 18.7% in 18 months, crossing the national benchmark
91% fault prediction accuracy before physical failure, up from 74% at deployment
₹47 Crore in revenue recovered in Year 1 through anti-theft drives and billing adjustments, independently audited
62% reduction in downtime; average consumer outage duration fell from ~6.8 hours to ~2.6 hours per incident
Metric : AT&C losses (network average) | Before Platform : 28.4% | After 18 Months : 18.7% (↓34%)
Metric : Fault prediction accuracy | Before Platform : N/A (reactive only) | After 18 Months : 91%
Metric : Mean time to fault detection | Before Platform : 4.2 hours (post-failure) | After 18 Months : 14 minutes (pre-failure)
Metric : Unplanned transformer failures | Before Platform : Avg. 18/month | After 18 Months : Avg. 5/month
Metric : Non-technical loss consumers identified | Before Platform : , | After 18 Months : 13,400+ confirmed
Metric : Consumer outage duration | Before Platform : ~6.8 hours/incident | After 18 Months : ~2.6 hours (↓62%)
Metric : Revenue recovered | Before Platform : , | After 18 Months : ₹47 Crore in Year 1
Behind these numbers is a more basic shift: the mean time to fault detection moved from 4.2 hours after a failure to 14 minutes before one. Unplanned transformer failures dropped from an average of 18 a month to 5. And in a network where theft had long been suspected but never quantified, more than 13,400 non-technical loss consumers were identified and confirmed.

Operational Impact

For consumers, this meant materially shorter outages and fewer unplanned interruptions. For field engineers, it meant dispatch based on evidence and asset condition rather than fixed calendars or complaint volume. For the control room, it meant visibility into network health before problems escalated, rather than after. As the utility's director summarised it, the return on investment crossed expectations within the first year, but the less quantifiable value was that the control room finally knew what was happening on the network before the phones started ringing.

Three design choices separated this platform from an off -the-shelf alternative and were largely responsible for driving real adoption rather than just deployment.

Built for Indian grid conditions. Monsoon-linked fault seasonality, rural-urban load heterogeneity, unmetered connections and open-access consumer dynamics were built into the core model logic from the outset, rather than handled as exceptions layered on top of a generic model.

No black boxes. Every AI output carries a plain-language justification, which meant field engineers could interrogate the system's reasoning rather than simply trust or distrust its output. That transparency is what allowed engineering judgment and AI recommendations to reinforce each other instead of competing.

End-to-end ownership. Kreate Technologies stayed embedded through the full 18-month engagement, refining models, fixing pipeline edge cases, training new operators, and iterating the interface based on real usage. A dedicated AI trainer sat in daily operations meetings for the first six months, which meant the platform kept improving in step with how the utility actually worked.

Key Takeaways

Data discipline has to come before modelling. Spending the first three months on a clean, unified data pipeline, before any model reached production, is what made the 74%-to-91% accuracy improvement possible later. Models
Explainability drives adoption more than raw accuracy does. Field engineers stopped second-guessing the platform once it started explaining its reasoning, not just its conclusions. A highly accurate model that operators don't trust delivers far less value than a transparent one they act on.
Purpose-fit modelling beats a single general model. Matching each problem to the right technique- LSTM for temporal fault patterns, XGBoost for asset health, Isolation Forest and DBSCAN for anomaly detection, Prophet-based forecasting for seasonal demand- produced better results than forcing one architecture to handle every use case.
Embedded delivery accelerates adoption. Having an AI trainer inside daily operations meetings for six months meant the platform was shaped by how the utility actually worked, not how a project plan assumed it would work. That proximity is what turned a deployed system into an adopted one.
A platform, not a project. With a second phase already underway, extending into substation-level automation and demand-response, the engagement was structured from the start as an evolving capability rather than a one-time delivery.

AI-Powered Grid Fault Intelligence: How a State Distribution Utility Cut AT&C Losses and Learned to See Its Own Network | Kreate Technologies | Kreate Technologies