How Cascade works,
and where it doesn’t.

Cascade is a planning sketch, not a forecast. It shows how failures could travel through a city’s infrastructure, given open map data and a set of assumptions you can see and change. Every number it shows is a simulated scenario.

Data

Infrastructure and roads: OpenStreetMap, via the Overpass API, for a 10 × 10 km area of Pune, Chennai, Valencia: substations, water works, pumping stations, water towers, hospitals, clinics, telecom masts, fuel stations, shelters, bridges, neighbourhoods, major roads, rivers, canals, coastline and power lines.

Elevation: AWS open Terrain Tiles (Terrarium encoding, zoom 13). These are surface models; buildings and trees can lift the ground they report.

Live hazard seed: Open-Meteo’s forecast API (72-hour rainfall) and Flood API (GloFAS river discharge). Optional; the simulator works without it.

Explanations: optional Gemini free tier, which only narrates a failure chain the engine already computed. Without a key, a fixed template says the same thing.

The model

Graph. Nodes are mapped facilities and neighbourhoods. Edges are dependencies of four kinds: power, water, hospital access and comms.

Hazard. A river flood as a bathtub model on height above the nearest mapped drainage (river, canal or coast). The stage rises linearly to its peak, then holds. A node fails when the water at it is deeper than its threshold; a road closes when it is deeper than the passable depth; a bridge fails above its deck clearance.

Cascade. A discrete-event simulation in 15-minute steps over 72 hours. A node keeps working while its inputs are met, or while the backup buffer for a lost input lasts (a generator, a battery, a tank), and fails the step the buffer runs out. AND edges must all be met; OR edges need any one supplier.

Road access. For each access dependency, the engine finds the route whose lowest point is highest (a bottleneck path). Access is lost when the water rises above that point.

Uncertainty. Monte Carlo over the peak stage and over which supplier an inferred edge really points at, giving the probability each node fails within 72 hours.

Optimizer. For a budget of up to three interventions, exhaustive search over the 15 best single moves. Above three, greedy search with lazy evaluation.

Tests. Unit tests cover a simple chain, AND versus OR logic, buffer expiry at the exact step, flood thresholds, bottleneck road access, that adding a backup never makes an outcome worse, determinism with a fixed seed and a golden replay.

Impact

Impact = Σ over neighbourhoods d, services s of ws × populationd × hoursd,s without service ÷ 1000

The result is thousands of criticality-weighted person-hours of essential service lost over 72 hours. Services are power, water, access to a working hospital, and comms. Weights and the population used where OSM has none are assumptions, listed below.

Provenance

  • observedTraced in OpenStreetMap: a mapped power line runs from the substation to the facility.
  • inferredGuessed from geography: the nearest supplier within a service radius, with a confidence. Most edges are inferred, because OSM rarely records what feeds what.
  • user-setConfirmed, re-weighted or removed by you in the dashboard.

Assumptions

Every one of these is a planning assumption, not a measurement. Each can be changed in the dashboard and results recompute.

Backup buffers

Hospital generator fuelOn-site diesel stock before a resupply is needed.48 h
Hospital water storageRooftop and sump tanks.24 h
Hospital fuel resupply windowHow long a hospital can go without a reachable working fuel station.24 h
Hospital comms fallbackLandline and radio fallback before coordination breaks down.12 h
Clinic backup powerSmall inverter or generator.4 h
Clinic water storage8 h
Mast batteryBattery bank at a mobile mast.4 h
Water tower storageHours of normal draw held in an elevated tank.8 h
Pumping station backup powerAssumes no on-site generator.0 h
Water works backup power2 h
Fuel station backup powerPumps stop when power does.0 h
Shelter backup power6 h
Shelter water storage12 h
Household water storageTypical overhead tank in a neighbourhood.6 h
Phone batteryBefore a district counts as without comms.2 h

Flood depth that fails a node

SubstationSwitchgear at ground level.0.3 m
Water works0.5 m
Pumping station0.3 m
Water towerElevated tank, but valves sit at ground level.1 m
HospitalGround-floor plant rooms.0.5 m
Clinic0.3 m
Telecom mastEquipment cabinet at the base.0.5 m
Fuel station0.3 m
Shelter0.5 m
Bridge deck clearanceHeight of a bridge deck above the ground under it.4 m
Passable road depthDeeper than this and a road counts as blocked.0.3 m

Dependency inference

Substation service radiusA consumer is fed by its nearest substation within this distance.5 km
Water source service radiusNearest tower first, else nearest pumping station or works.6 km
Mast coverage radiusAny mast in range can serve (OR logic).2.5 km
Masts counted per consumer3
Hospitals counted per districtA district has care if any of these is up and reachable.2
Fuel stations counted per hospital2

Criticality weights

Power1 x
Water1 x
Access to a hospitalWeighted highest: losing it is the most direct risk to life.1.5 x
Comms0.5 x
District population when OSM has noneUsed only where the OSM place has no population tag.20000 people

Hazard and simulation

Time to peak stageWater rises linearly to the peak, then holds.6 h
Stage uncertainty (1 sigma)Spread of the peak stage across Monte Carlo runs.0.3 m
Live seed: metres per doubling of dischargeTurns a forecast discharge ratio into a stage. A rough rating curve, not a hydraulic model.1.5 m
Monte Carlo runs200

Interventions

Added generator runtime72 h
Raise or harden byFlood walls, raised plinth or relocated switchgear.1 m
Mobile substation arrives after12 h

Limits

  • The bathtub model ignores flow, timing along the river, drainage capacity, embankments and rainfall run-off. Height above the nearest drainage is a straight-line approximation, not a flow path.
  • The elevation data is coarse (tens of metres) and is a surface model, so river banks and built-up ground can be misread by metres.
  • Most dependencies are inferred. Real feeder networks, water mains and mast backhaul are not public. Treat every inferred edge as a hypothesis.
  • Backup durations, thresholds and weights are generic defaults, not facility surveys.
  • OSM coverage varies: Pune has 8 mapped substations and 2 telecom masts in the area, so single substations look more critical than they may be.
  • Restoration is modelled only for the mobile-substation intervention; repair crews and failover switching are not modelled.

Credits

  • Map data © OpenStreetMap contributors (ODbL). Basemap tiles: OpenFreeMap.
  • Elevation: AWS open Terrain Tiles (Mapzen terrain, derived from SRTM and other public sources).
  • Landing footage: “Obere Schleuse in Thun” by Nachtbold, CC0, via Wikimedia Commons.
  • Sky footage: “2011-08-05 21-45-00-timelapse-clouds-seloncourt” by Thomas Bresson, CC BY 3.0, via Wikimedia Commons; cropped, colour graded and looped.
  • Weather and river discharge: Open-Meteo (CC BY 4.0), GloFAS via the Copernicus Emergency Management Service.
  • Explainer video: footage by Marsupium (CC0), Nachtbold (CC0) and “City skyline (time lapse)” (CC BY 3.0) via Wikimedia Commons; music “Ambient Documentary Build up” by tyops (CC BY 4.0) and sound effects by Joao_Janz (CC0) via Freesound; voiceover generated with Gnani Vachana. Full list in CREDITS.md.

Open the simulator