Flood detention area digital twin: the forecast-to-plan relay across 2,316 km²

An auxiliary decision platform for a flood detention area on the lower Yellow River: four counties, 45 townships, 1,600 villages and 1.3171 million people live inside 2,316 km², with a diversion capacity of 10,000 m³/s. Six menus — home, monitoring, forecasting, warning, rehearsal, response plan — are strung into one timeline that runs from “the river is rising” to “everyone is out”.

Industry
Water resources · flood control / detention area operation
System type
Digital-twin decision platform (GIS one-map + forecast/warning/rehearsal/plan + sluice BIM + inundation rehearsal)
Key technology
GIS layers and inundation polygons · discharge-to-stage conversion · rehearsal parameter configuration · graded warning with SMS broadcast · evacuation ledger · 3D sluice models · multi-source integration
Jiding's role
Requirements and menu restructuring, prototypes and UI, front-end development, GIS and data integration, rehearsal parameters and result presentation, launch support
2,316 km²Total detention area, about 140 km long between two levees
20 bn m³Effective storage; the diversion sluice is designed for 10,000 m³/s
1600Villages inside, across 45 townships and 1.3171 million people
56Gates on the diversion sluice; the outlet sluice has 6 and can backflow

Why It Is Hard

The problem is not the absence of a platform. It is that nobody dares press the button.

A detention area is the last card in the flood-control hand: on ordinary days it is home to 1,600 villages, and in an extreme flood it is 2 billion cubic metres of storage. The card may be played once in decades, and the price of playing it is hundreds of thousands of people leaving overnight. So the decision-maker faces a brutal trade: not using it risks the downstream; using it may flood the area for nothing. The platform exists to close that information gap.

The base is unclear: how many people, how many villages?Population moves, villages merge, farmland and factories change every year. The paper ledger and the ground truth diverge, and an evacuation plan built on it will miss.
Put people into the layer first.Counties, townships, villages, population, farmland and storage volume are entered once and aggregated two ways — by county and by inundation zone. The home page states 1,600 villages and 1.3171 million people, and every report points back to that one base.
The threshold is vague: how much water before diversion?Upstream discharge changes continuously; diversion is a binary act. The threshold lives in the plan, and the plan is a thick book nobody opens on site.
Turn the threshold into a choice.Discharge levels from 5,000 to 22,000 m³/s are listed with the stage at each section and the matching plan. Pick a level and stage, inundation extent and evacuation scale move together — no more flipping through paper.
The consequence is invisible: after diversion, where and how deep?Inundation maps are usually static results. Change the flow or the gate pattern and the model has to run again — by the time results arrive, the decision window has closed.
Make the rehearsal adjustable.Diversion flow, duration, number of gates and opening height, backflow through the outlet: move any of the four and the four result cards — inundated area, affected people, economic loss, stored volume — move with them.
Nobody is reached: who actually got the warning?Most residents install no app and read no website, and phones change numbers during flood season. “Published” in a system does not mean anyone received it.
Select by base station, not by account.Draw the area on the map, tick the carriers’ base stations covering it, and the carrier broadcasts an SMS to every phone inside — about 586 km² and 100,000 people per station, with the reach counted before sending.
The loop does not close: after the warning, who leaves, by which road, to where?Routes, shelters and responsible officers sit in different files; on site it is phone call after phone call, and the more urgent it gets the messier it is.
Hang the ledger on the map.Fourteen evacuation records list origin, shelter, headcount and route, with shelters 1 and 2 pinned on the map — names, routes and points on one screen.
No record for review: was the call right?The process scatters across meeting notes, calls and paper orders. Afterwards nobody can reconstruct what the operator was actually looking at.
Leave the process as data.Monitoring curves, warning records, rehearsal parameters and plan versions sit on one timeline, so a review replays the event instead of relying on memory.

Policy background: the Ministry of Water Resources guidance on digital twins for reservoirs, sluices and detention areas (2024, No. 269) sets 2027 for advancing national detention-area digital twins and 2030 for completion; the 2024 flood-defence work plan likewise encourages “digital mapping, intelligent simulation and forward-looking rehearsal”. The six-menu structure follows that forecast-warning-rehearsal-plan framework.

The Area

See the ground first: a long narrow bowl with 1.3 million people in it

A detention area is not a reservoir. It is a valley floor pinched between two levees: the south levee is the Yellow River dike, the north levee is a historical embankment, and between them the strip is at most a dozen kilometres wide and about 140 km long. When it operates, water enters through the upstream sluice, spreads from the low ground outward, and finally leaves through the downstream outlet. Only with that bowl in mind does it make sense that the same discharge level means completely different risk in different villages.

Design storage level South levee (Yellow River dike) North levee ShallowDeepMain flowBasin Shallow zone231 villages · 163.9k Deep zone658 villages · 528.4k Main flow327 villages · 313.6k Basin384 villages · 320.2k Diversion 56 gates10,000 m³/s Outlet 6 gates1000 m³/s Schematic section, vertical scale exaggerated; zone populations from platform demo data
The same flood line is ankle-deep in the shallow zone and roof-deep in the deep zone. That is why risk cannot be averaged over total area.

Four zones, four fates

The platform cuts the area into shallow, deep, main-flow and basin zones, each carrying a village count and a population. This is not decoration: evacuation order, shelter assignment and compensation all follow the zone. The deep zone — 658 villages, 528,400 people — is the bulk of the evacuation; the shallow zone has fewer people but dries out first, so it is usually the first cohort to return.

Putting the four zones in one table has a second effect: it forces the data to reconcile. The four must add up to 1,600 villages and 1.3171 million people, and one missing village sends someone back to check. The table is the reconciliation sheet.

One table, four counts

ZoneVillagesPeople (000)
Shallow231163.9
Deep658528.4
Main flow327313.6
Basin384320.2
Total1,6001,317.1

If the four do not add up, the base is wrong. Reconciliation is routine, not a pre-flood scramble.

The Skeleton

Six menus are not six screens. They are a relay.

The menu structure follows the upper-level integrated platform: home, monitoring, forecasting, warning, rehearsal, plan. At first glance they look like six parallel modules; in use they are one timeline — watch the forecast as the river rises, warn as it nears the threshold, rehearse before deciding, execute the plan once the button is pressed, and monitor throughout. Drawn as six boxes, users treat them as six pages; drawn as a line, they know which one to open when.

HomeOne map
Station counts
Profile and zones
Roads and bridges
MonitoringArea / people
Economic loss
Stored volume
Rain, stage, video
ForecastingDischarge
Stage
Rain and temperature
Structure stages
WarningOverview
Base-station broadcast
Supplies and hazards
Rain and works
RehearsalParameter setup
Gates and heights
Inundation run
Scenario compare
PlanMatch by discharge
Organisation
Evacuation
Gate operation
T-48hT-24hT-12hT-6hT-0 Monitor rain, stage, flow video, patrols Forecast 48h rain and flow stage per section Warn yellow / orange / red SMS by base station Rehearse how many gates, how high where it floods, who is hit Plan / operate who leaves, which road drainage and return Decision window: 24–12 h Evacuation: 12–6 h Execution: 6–0 h The further right, the fewer the options and the higher the cost. The platform exists to push decisions left. Operating early without reaching the threshold is also a misjudgement — both ends need evidence.
A relay: forecasting buys time, warning buys consent, rehearsal buys evidence, the plan buys order. Drop one baton and the rest cannot make it up.

Two pages sit outside the relay: the 3D models of the diversion and outlet sluices. They are the “see the site” entry used during rehearsal — gate opening sequence, flow behaviour, chamber structure — so that inundation results and gate structure are not crammed into one picture.

Home: One Map

The home page must answer one question: if we divert now, who gets flooded?

A 1920×1080 wall: the area as one map, with the four zones tinted over terrain imagery; six station counts across the top; weather and structure stages on the left; profile, zones and evacuation roads on the right. It looks full, but it serves one question — if diversion starts now, how big is the extent and how many people are inside it.

20Hydrological stations, including four key control stations upstream
50Rain gauges across the area and the upstream catchment
150Video channels at sluices and vulnerable levee sections
20Defence priorities; plus 15 response teams and 15 obstruction structures
Home page of the detention area platform: one central map with station counts, profile and zones de-identified
Home: the six counts are the inventory, the map is the space, the three panels are the base. One coordinate system, so every panel can drive the map.

Why the top row shows counts, not icons

The first version dropped hydrological, rain and camera icons straight onto the map. Two hundred and twenty icons covered the map and hid the inundation extent. The fix: stations move into monitoring and are plotted on demand; the home page keeps the counts. Counts answer “how much do we have”, the map answers “where does the water go”. Separated, both stay legible.

The three right-hand panels reconcile

The profile gives four counties, 45 townships, 1,600 villages and 1.3171 million people; the zone panel gives villages and people per zone; the road panel gives 129 evacuation roads, 897.5 km and 75 bridges. The three interlock, so updating one forces a check on the other two — data governance built into the interface.

The home page also keeps a toolbox (distance, area, coordinates) and a layer switch. Distance is a high-frequency need in flood season: when the field reports “800 m east of that village”, measuring it on the map halves the conversation.

Monitoring

Monitoring does not end at a curve. It ends at a list you can phone.

Four cards sit on top: inundated area, affected people, economic loss, stored volume. Below them, a fourteen-row evacuation ledger: origin, shelter, headcount, route. The top is “how much”; the bottom is “who”. Only the top is for briefings; both together are for the field.

Operation monitoring page with four statistic cards, an evacuation ledger and the inundation map de-identified
Operation monitoring: four cards give magnitude (589.6 km², 156,000 people, CNY 158.9 million, 1.256 billion m³), the ledger gives destinations row by row, the map gives shelter positions.

Four cards, four counter-intuitive points

A larger inundated area is not automatically worse.589.6 km² sounds alarming, but if the water sits mostly in the basin and deep zones the affected population can be far below a uniform estimate. Area must be read with zones.

Loss needs its basis stated.CNY 158.9 million at what unit price, including crops or not, counting shutdowns or not? If the card does not say, the review will argue. The basis is a configuration item.

Storage and area are not linear.1.256 billion m³ over 589.6 km² averages about two metres — but the deep zone may be five and the shallow zone under one. Averages mislead.

Why the ledger needs a route column

“2,500 people from that village” is useless on its own. The field needs the origin, the road, the shelter and the counterpart. The route column is what turns a number into an action.

All fourteen records locate on the map: origin, destination, connecting line. Shelters share one icon, and shelters 1 and 2 stay pinned so nobody has to ask which one still has space.

Forecasting

The translator: turning a discharge level into how deep it gets at which village

A hydrologist says “18,000 at the control station”, and the field hears a number; the village head wants to know where the water stops. The whole job of the forecasting page is to translate discharge into stage and stage into inundation. The translation comes from two places: the stage–discharge relation at each section, and the plans written per discharge level.

Stage forecast page: a table of peak stage and maximum depth per section, with structures marked on the map de-identified
Stage forecast: pick an upstream discharge level and the left table lists peak stage and maximum depth per section, with the corresponding structures marked on the map.
Upstream discharge levelStage at the four gauging stationsWhat the platform prompts
5,000–10,000 m³/sBelow warning; stages rising section by sectionDenser monitoring; no evacuation
10,000–15,000 m³/sApproaching warning levelYellow: preparation
15,000–20,000 m³/sSome sections above warningOrange: evacuation and clearance
20,000–22,000 m³/s and aboveNear the guaranteed stageRed: detention-area operation; move into rehearsal

22,000 m³/s is the design flood standard of the lower Yellow River levees, and the key node for whether a detention area has to be used: only when reservoir regulation and existing flood-diversion works still cannot keep the downstream within safe discharge does the operating procedure begin. The platform turns that judgement into an option instead of a number to memorise.

Warning

Three levels, six base stations: warn the phone, not the account

Three levels map to three stages of action: yellow is preparation, orange is evacuation and clearance, red is operation. The level is not decoration; it decides who moves and when. The hard part is the last mile: reaching people who have installed nothing.

YellowIII
TriggerUpstream discharge nears the warning threshold and stages are forecast to keep rising
ActionFlood teams on duty, supplies counted, boats moored; notice to cadres and duty officers, without alarming residents
OrangeII
TriggerDischarge keeps climbing, some sections above warning, diversion becomes likely
ActionEvacuation begins: move people by zone and by ledger, open shelters, control roads
RedI
TriggerNear guaranteed stage; operation approved
ActionBroadcast to the whole area by base station; execute the diversion plan, open gates in sequence, verify village by village
4 5 3 6 Station 1 Station 2 Select by station, count first, send second Tick stations → overlay coverage → estimate reach → apply template Coverage per station~586 km² People per station~103,500 Combined estimated reach13,500 Message templateby level and area Dispatch and receiptlogged by batch Coverage varies with radius and terrain; the screen shows an estimate Actual reach follows the carrier receipt; sending needs a second confirm Schematic: hexagons are base-station cells; red dashed rings are the estimated coverage of selected stations
“Select people by base station” is the least software-like trick in the system and the most effective: no sign-up, no network, just a phone inside the ring.
Warning dispatch page: tick base stations, see estimated reach, edit the SMS template de-identified
Warning dispatch: tick stations 1–6, see the estimated reach at the top (13,500), edit the template, press send; each station on the map carries its coverage area and population.

Rehearsal

Fifty-six gates cannot be one “open” button

The left side is a set of parameters: diversion flow, duration, number of gates and opening height on the diversion sluice, the same for the six outlet gates, and whether backflow is allowed. The right side is four results: inundated area, affected people, economic loss, stored volume. Parameters are “how the water is released”, results are “what happens next”; the model sits in between.

Diversion sluice · 56 gates (14 × 4) ONONON Example: three gates open at a 3 m opening; the sequence follows the plan and is staged, not all at once Each gate is 10 m wide and 7.4 m high; all 56 open equals the design capacity of 10,000 m³/s Outlet sluice · 6 gates ON Outlet 1,000 m³/s Backflow up to 1,000 m³/s Drains internal water at 270 m³/s Backflow is a separate parameter: the same river, water going the other way, a different outcome.
Opening a sluice becomes 56 + 6 clickable cells: real operation is staged, gate by gate and hour by hour. With a single master switch, the rehearsal would only ever have one answer.
Diversion rehearsal page: parameters on the left, four result cards on top, inundation map shaded by depth de-identified
Diversion rehearsal: parameters on the left (flow, duration, gates and heights, backflow), results on top (589.6 km² / 156,000 people / CNY 158.9 million / 1.256 billion m³), map shaded by depth.

A rehearsal result is not a forecast; it is a run under given parameters. Every result card carries its parameter snapshot, so a second run is a second snapshot — and two side by side are a scenario comparison. That is why the rehearsal has to be adjustable: a fixed run cannot support the decision of which option to choose.

Response Plan

The plan: a thick book, split into five clickable blocks

The paper plan is not badly written; it is hard to find things in when you need them. The platform organises five blocks: area profile, organisation and support, warning and alarm, evacuation and shelter, sluice operation. Each block maps to one action on site, not to a chapter in a document.

Evacuation points Shelters Point 1 · 2,500 people Point 2 · 1,500 people Point 3 · 900 people Point 4 · 12,500 people Point 5 · 3,500 people Point 6 · 22,200 people Shelter 1 · education park Shelter 2 · sports centre Shelter 3 · township school Line width follows headcount; in the platform each line opens to show route, owner, vehicles and phone numbers Points are numbered for de-identification; names are withheld
“How many people go where” drawn as lines: clickable, readable, weighted. That is all a command post needs.
Plan page: station stage table and evacuation ledger on the left, zone and road statistics on the right de-identified
Plan page: gauge stages and the evacuation ledger on the left, zone statistics and evacuation roads on the right; four entries along the bottom open the profile, organisation, warning and evacuation blocks.

One easily missed design: the plan page is selected by discharge level. The same screen at 15,000–20,000 and at 20,000-plus produces different evacuation scale, road closures and shelter counts. A plan is not a document; it is a set of branches triggered by conditions.

Iteration

20 change requests and 13 open questions: what the schedule actually looked like

Within one month of kick-off the operating authority filed 20 change requests, all marked urgent: ten on the home page, one on monitoring, one on forecasting, three on warning, one on the plan, one on rehearsal and two for the sluice models. This is not rework; it is requirements converging against real data. Read the 20 and the current design explains itself.

Where the changes concentratedTypical itemThe real need behind it
Home page (10)Add detailed geometry to the boundary and draw it blue; default to the four upstream stations; move station icons to monitoringThe home page is not “everything”; it is boundary plus four key stations first
Home interactions (4)Profile, zones, structure stages and road statistics must all “interact with the map once data arrives”Every panel has to point back to the map; no dead images
Monitoring / forecasting (2)Follow the floodplain platform’s layers and plotting; align the forecast page’s five modulesThe upper-level platform is the house style; the language must match
Warning (3)Supplies warning to the materials platform; hazard warning from smart sensors; a new alarm buttonWarning is not rain alone; it covers supplies and structural hazards
3D models (2)Diversion sluice by 27 April, outlet sluice by 10 MayRehearsal must “see the gates”; the models are hard milestones

The 13 open questions matter more than the 20 changes

The communication list holds 13 unresolved items: map elevation accuracy and data format, model level for the sluices (down to parts?), oblique photography versus design models, which rivers to display, who supplies the two expressways and village data, which system provides structure records, the sensor inventory, the coverage of weather forecasts, services versus WebAPI, deployment at the provincial platform or locally, who provides the inundation model, and whether a test environment exists.

These are boundary questions, not technical ones. Unclear boundaries can waste any amount of work. Our approach: list them, build what does not depend on them (skeleton, interaction, presentation, parameter configuration), make dependencies configurable, and wire them when data lands.

Why the two sluices came first

The platform has six menus and dozens of screens, but the hardest and most memorable are the two 3D sluice pages: opening sequence, flow behaviour, chamber structure. Doing them first buys three things: it proves the 3D pipeline and performance, it gives decision-makers a visible anchor of trust, and it forces the data owners to hand over models and drawings early.

It is the same order we use on industrial and water projects: take the hardest block first, never leave the risk to the last week.

One honest note: acceptance for this kind of platform is rarely “every line ticked”. It is whether, during a pre-flood drill, the duty officer can find the map, the person and the gate he needs within three minutes. So before handover we run a “find it” test, not just a feature demo.

Delivery & FAQ

What is delivered, and five questions we get asked

Delivery covers the platform, data integration, the two sluice model pages, deployment and training, plus a maintainable statement of data definitions. The definitions matter more than the code: where each number comes from, how often it updates, and who owns it.

Q1How is a detention-area twin different from a city twin?
A city twin manages static assets and live status; a detention-area twin manages “what happens if we release water”. Its core is a chain that has to compute end to end: upstream discharge → section stage → diversion flow and duration → inundation extent and depth → affected villages and people → routes and shelters. If any link is guessed, everything downstream is void.
Q2Where do the inundation data and models come from?
Three sources: hydrological and hydrodynamic models plus inundation algorithms from the provincial or basin authority; existing flood-risk maps and historical event results; local oblique photography and terrain. The front end configures parameters, calls the model and presents results — it does not invent the model. Who provides it, how it is called and in what format goes on the open-questions list at kick-off.
Q3Why warn through carrier base stations?
Because the audience is not “users” but people in 1,600 villages who have installed nothing. Select the area by base station and the carrier broadcasts an SMS to every phone inside — about 586 km² and 100,000 people per station. No install, no login. It is the hardest last-mile channel available; the trade-off is text only, no interaction, so graded templates and dispatch logs are mandatory.
Q4How do you “open” a 56-gate sluice on screen?
No single master button. The rehearsal page breaks the 56 gates into a clickable matrix, each with open/closed state and opening height; the six outlet gates form a separate group with a backflow switch. Parameters then drive the model. Real operation is staged by gate and by hour; with only a master switch, every run would produce the same answer.
Q5Where do projects like this usually stall?
Data boundaries, not technology. Who supplies coordinates, who maintains population, what the interface looks like, whether there is a test environment, where it is deployed — all “to be confirmed”. List them explicitly, build what does not depend on them, keep dependencies configurable rather than hard-coded, and no screen has to be rebuilt.

MORE CASES

More Related Cases

Flood detention area digital twin
Digital TwinSmart flood prevention, intelligent flood control

Flood detention area digital twin

An auxiliary decision platform for a flood detention area on the lower Yellow River: four counties, ...

View Details →
Campus PV energy platform
Digital TwinEnergy consumption management, clean energy

Campus PV energy platform

An integrated energy platform for an aviation-base campus, delivered as a 3D digital twin plus a 2D ...

View Details →
Digital-twin bird-repellent system
Digital TwinEcological bird repelling, visual analysis

Digital-twin bird-repellent system

A bird-hazard platform for a regional grid operator's transmission maintenance team, delivered as a ...

View Details →

View All Cases →

Planning a similar production line or enterprise system?

Start with our pricing logic, then let's talk requirements — cost structure, technology comparison and acceptance criteria, all laid out in the open.