Grid Resilience Operations Center

An energy-vertical demo showcasing NVIDIA's AI stack running on Red Hat OpenShift. The demo simulates an ice storm scenario on a distribution corridor, demonstrating AI-powered weather forecasting, automated defect detection, GPU-optimized crew dispatch, and a real-time 3D substation digital twin.

Note

This project was developed with assistance from AI tools.

What It Does

A scripted demo walks through storm preparedness and response for an electric utility operations center:

• Weather forecasting with NVIDIA Earth-2 CorrDiff for hyperlocal 3km downscaling
• Camera-based defect detection with Cosmos Reason for AI observation of fixed infrastructure cameras
• Crew dispatch optimization with cuOpt for GPU-accelerated vehicle routing
• 3D substation digital twin with Omniverse Kit streaming via WebRTC
• Risk scoring across weather exposure, asset age, vegetation proximity, and inspection recency

The console provides a map view with weather overlays, asset risk rankings, camera findings, crew route visualization, and customer impact tracking through each phase of the scenario.

Cluster Prerequisites

• OpenShift 4.x cluster
• NVIDIA GPU Operator installed
• 3x L40S GPU nodes (CorrDiff, Cosmos Reason via vLLM, Omniverse Kit)
• 1x L4 GPU node (cuOpt)
• AMQ Streams or Strimzi Kafka operator
• Crunchy PGO (PostgreSQL operator)
• Argo CD for GitOps deployment

Running

Scale up

oc scale deployment/corrdiff-nim deployment/vllm-cosmos deployment/cuopt deployment/kit-substation deployment/coturn -n grid-ops-ai --replicas=1
oc scale deployment/weather-service deployment/scenario-engine deployment/camera-simulator deployment/defect-detector deployment/dispatch-optimizer deployment/risk-engine -n grid-ops-services --replicas=1
oc scale deployment/console-backend deployment/console-frontend -n grid-ops-console --replicas=1

Scale down

oc scale deployment/corrdiff-nim deployment/vllm-cosmos deployment/cuopt deployment/kit-substation deployment/coturn -n grid-ops-ai --replicas=0
oc scale deployment/weather-service deployment/scenario-engine deployment/camera-simulator deployment/defect-detector deployment/dispatch-optimizer deployment/risk-engine -n grid-ops-services --replicas=0
oc scale deployment/console-backend deployment/console-frontend -n grid-ops-console --replicas=0

Seed the database

oc port-forward -n grid-ops-infra pod/<postgres-pod> 5432:5432 &
python3 data/seed/load_seed_data.py "postgresql://gridops:<password>@localhost:5432/gridops?sslmode=require"

Seed the CorrDiff cache

After the weather-service and CorrDiff NIM are running:

oc port-forward -n grid-ops-services svc/weather-service 8180:8080 &
curl -X POST http://localhost:8180/forecast/run

This runs a live CorrDiff inference (~45s) and caches the result. Subsequent scenario starts use the cache for instant replay.

Deploy the console

The console frontend requires cluster-specific values for Kit streaming and TURN relay:

CLUSTER_APPS=$(oc get ingresses.config cluster -o jsonpath='{.spec.domain}')

helm install console-frontend infrastructure/helm/console-frontend/ -n grid-ops-console \
  --set kitSignalingServer=kit-substation-grid-ops-ai.${CLUSTER_APPS} \
  --set turn.urls="turns:turn-grid-ops-ai.${CLUSTER_APPS}:443?transport=tcp" \
  --set turn.credential=<coturn-password>

Access the console

oc get route console-frontend -n grid-ops-console -o jsonpath='{.spec.host}'