Gen AI · evidence-led · Dhofar utilities

Where Gen AI is actually useful on Dhofar's grid and water network.

A practical filter, not a sales pitch. Eight use-cases where the evidence supports deployment, four where it does not, and the engineering principles that separate the two.

Browse simulators See the data

Production models

Top accuracy

96.4%

Runs / day

1,284

Use-cases worth piloting

Areas not ready

+31%

PM gain (published study)

Human-in-loop

On every safety-critical action

A note on tone

Gen AI is a tool, not a strategy. Most published utility wins come from a narrow band of patterns: synthetic data for rare events, semantic retrieval over operational text, and scenario generation. Everything else is either pre-Gen-AI ML in a new wrapper, or a pilot that has not survived its first regulatory review.

What to build

Use-cases with evidence and a worked example

Each entry states the problem, the Gen AI mechanism, a concrete Dhofar example, the evidence and the caveat that determines whether it survives a pilot.

Asset maintenance

Predictive maintenance with synthetic failure data

Proven

Problem: Real failure records on Dhofar feeders, transformers and pumps are scarce. Classifiers miss the rare faults that cause outages and supply interruptions.
What: Tabular GenAI (VAE / GAN / diffusion) synthesises minority-class failure samples; train RUL and fault classifiers on the augmented set.
Example: Salalah 33 kV transformer: DGA pattern matched to incipient inter-turn fault, replacement scheduled in a 6-day Khareef window instead of a forced outage.
Evidence: +31% F1 vs. classical ML on published asset datasets. Industry baselines: +10–20% uptime, 5–10% lower O&M cost.
Caveat: Synthetic data must be validated against held-out real failures, otherwise the model learns the generator's bias.

Operations support

LLM-assisted work-order drafting & retrieval

Proven

Problem: Decades of unstructured text, work orders, NRW investigations, vendor manuals, that crews cannot search semantically during an incident.
What: Fine-tune an LLM on internal corpora; retrieve the closest historical case for a live alarm or leak; draft the work order, parts list and crew brief.
Example: Wilayat of Taqah: 'low pressure complaint' auto-matched to a 2023 PRV failure pattern; work order, valve SKU and crew assigned in under 90 seconds.
Evidence: Mature pattern in industrial RAG deployments, the primary near-term GenAI lever for guided-optimisation use-cases.
Caveat: Hallucination on safety-critical fields (ratings, settings, chemical doses) is unacceptable. Keep humans on every approval.

Water network

Leak & non-revenue-water detection

Proven

Problem: Non-revenue water hides in pressure noise and DMA balance gaps. Manual auditing is slow and reactive.
What: Sequence models learn DMA pressure/flow signatures and flag anomalies; generative scenarios simulate burst patterns to pre-train detectors.
Example: Salalah DMA-14: 11 l/s overnight excess flagged as a probable service-line burst; field crew confirmed within 4 hours.
Evidence: Operator deployments show 15–25% NRW reduction over 18 months when paired with active pressure management.
Caveat: Garbage-in: bad DMA boundaries or missing meter reads will dominate any model gain. Fix metrology first.

Renewables forecasting

Solar & Khareef wind generation forecasting

Promising

Problem: Volatile renewables on the Najd plateau and coastal corridor raise dispatch cost and ancillary-service procurement.
What: Sequence models (hybrid physics + transformer) trained on weather, telemetry and synthetic extreme-event scenarios (dust storms, Khareef onset).
Example: Day-ahead PV forecast for a 200 MW Najd plant: Khareef-monsoon scenarios generated synthetically because real instances in the window are too few.
Evidence: Established literature; GenAI's specific contribution is rare-weather augmentation, not the base forecaster.
Caveat: Marginal gain over a well-tuned classical forecaster is often single-digit percent. Build the boring baseline first.

Grid planning

Dynamic line & transformer thermal capacity

Promising

Problem: Lines are rated conservatively. Real thermal headroom changes with weather, load and asset condition, and is left on the table.
What: Forecast conductor and top-oil temperatures conditional on weather; propose dynamic ratings; generate synthetic heat-wave / Khareef scenarios for stress tests.
Example: Salalah – Thumrait 132 kV: model recommends a 5% temporary uprate during a forecast cool window, deferring a reconductor project by 18 months.
Evidence: Reduces congestion cost and capex deferral; impact is utility-specific and must be back-tested on local data first.
Caveat: Requires accurate asset records and IEC 60076 / IEEE 738 thermal models. Without them you are guessing with extra steps.

Customer & demand

Self-diagnosing customer complaints

Proven

Problem: Call-centre triage is slow; many complaints are explainable from telemetry already in MDM, SCADA and the GIS.
What: LLM cross-references the account, AMI interval data, outage history and weather; classifies the likely cause; drafts the response.
Example: 'Power flickered last night' → matched to a recloser operation 2.3 km upstream at 23:14; auto-response with explanation and ETA.
Evidence: Reduces average handling time and field roll-outs; deployable today on top of an existing CRM stack.
Caveat: Privacy: smart-meter and billing data are sensitive. Differential privacy and access controls are non-negotiable.

Wastewater operations

Treatment-plant influent & dosing forecasting

Promising

Problem: Khareef rainfall and tourism spikes swing influent load. Static dosing wastes chemicals or risks effluent breaches.
What: Probabilistic forecasts of influent flow and BOD/COD; generative what-if scenarios for storm events feed the chemical-dosing optimiser.
Example: Salalah East STP: 12% reduction in coagulant use over a Khareef month with no effluent quality regression.
Evidence: Well-suited to GenAI's scenario-generation strength; pilot evidence growing in Gulf utility deployments.
Caveat: Effluent compliance is the hard constraint. The optimiser proposes; the plant operator decides.

Operator training

Synthetic scenarios for control-room training

Promising

Problem: New operators lack exposure to rare contingencies. Real drills are infrequent and expensive.
What: Generate physically-plausible disturbance scenarios (cascading trips, dust storms, cyber events) with varied severity for simulator-based training.
Example: 30 synthetic N-2 contingencies seeded from one historical Salalah event; trainee performance scored against expert traces.
Evidence: Directly addresses the workforce-readiness gap noted across regional utility training programs.
Caveat: Scenarios must be vetted by senior operators; otherwise trainees learn artefacts of the generator.

Where Gen AI is not ready

Autonomous closed-loop control of safety-critical assets

No regulator currently accepts an LLM in the trip path. Keep humans on every protective action, electrical, hydraulic or chemical.

Free-form chatbots in the control room

Hallucination risk under stress. Constrain to retrieval, structured outputs and explicit confirmations.

Replacing physics-based planning tools

GenAI augments inputs and explains outputs. The hydraulic, power-flow and dispatch optimisation cores stay where they are, for now.

Cybersecurity automation without segmentation

GenAI in OT widens the attack surface. Defence-in-depth and zero-trust are prerequisites, not options.

Engineering principles

Six rules that separate pilots that ship from pilots that don't

Build the boring baseline first

Most 'AI gains' vanish when compared to a well-tuned classical baseline. Measure that baseline before any GenAI commitment.

Validate synthetic data against real holdouts

Augmentation is only useful if it improves performance on real, unseen failure or leak events.

Human-in-the-loop on every consequential action

Operators approve work orders, switching, valve moves, chemical doses, model proposes, human decides.

Treat data quality as the first-order problem

Asset records, DMA boundaries, AMI coverage and outage tags dominate model performance more than architecture choices.

Measure value in operational currency

Minutes of SAIDI, m³ of NRW saved, OMR of deferred capex, not BLEU scores or model size.

Constrain outputs to controlled vocabularies

Tag schemas, asset IDs, dose ranges and switching orders should be enums, not free text.

Model registry

Models running today across electricity, water and wastewater

Models

In production

Avg. accuracy

90.7%

Total runs

16,876

production

Electricity

Demand Forecast v4

Probabilistic load forecasting at feeder level using temperature, irradiance and behavioural priors.