In the autumn of 2025 the Han River (汉江) rose into numbered flood after numbered flood — seven in one season, by the Ministry of Water Resources' own count. Behind each decision about which reservoir gates to open, engineers on the Yangtze basin ran simulations: more than 80 rounds across the season, according to the ministry's account of it, replaying the same storm against different gate schedules. This was not a chatbot being asked a question. It was a river that exists twice — once in a riverbed, once in a server.
That second river is what Chinese water officials call a digital twin (数字孪生). Not a metaphor. It is a working model of a basin — channel geometry, reservoir operating rules, embankment elevations, soil moisture, live gauge readings — that can be pushed forward in time faster than the actual water moves.
The ministry put numbers on the effort at a December 2025 briefing and again in an October 2026 written reply to a policy advisory body. The headline claim: for major rivers, flood forecast lead time has gone from three days to ten, with accuracy in the critical period around 90%.
A river that exists twice
A digital twin basin is assembled from four layers, and only the last one is fashionable.
- Geometry. River cross-sections, floodplain terrain, dam and levee elevations, the storage curve of every reservoir.
- Rules. How each gate is operated, what downstream channel capacity is, which tributary must be protected first.
- Physics. Rainfall-runoff and hydrodynamic models that convert a forecast storm into a travelling wave of water.
- Data. Live inflow, level and rainfall readings that keep the model pinned to reality instead of drifting.
The models run as ensembles: dozens of variants of the same storm, some wetter, some nudged tens of kilometres east. The output is a spread rather than an answer — a fan of possible peak levels at named stations, and gate schedules that perform acceptably across the whole fan.
Worth stating plainly, because "AI" has become a catch-all: most of the heavy lifting is done by calibrated physical models, not by a large model (大模型) of the kind that writes essays. Machine learning shows up where physics is weakest — correcting bias in rainfall forecasts, filling gaps where gauges are sparse, speeding up simulations.
What the ministry says it built
The build-out, per the ministry:
- 94 pilot digital-twin basin tasks completed across 56 organisations, and a national water data foundation at petabyte scale holding 72 categories and about 26 million object records.
- A monitoring network of 34 remote-sensing satellites, more than 180 rain radars, 136,000 ground hydrological stations and over 300,000 engineering safety monitoring points.
- Eleven key projects in operational use, including the Three Gorges (三峡), the South-North Water Transfer (南水北调), Xiaolangdi (小浪底), Danjiangkou (丹江口) and Datengxia (大藤峡).
- A further wave under construction: 12 smart dams and 98 digital-twin flood detention areas (蓄滞洪区).
The ministry calls the network "sky, air, ground, water, engineering" (天空地水工) and treats it as the foundation for everything above it.
The three numbers that carry the argument
Three figures do most of the work.
2.72 million station-level flood forecasts (272万站次) issued since the start of the 14th Five-Year Plan period (2021-2025).
Three days to ten days. The forecast horizon for major rivers, the number that determines whether a city has time to move people and pre-release reservoir storage.
90%. Accuracy of major-river forecasts during critical periods, as stated by the ministry. The same briefing reports 12,400 flood warnings, 1.21 million flash-flood warnings, and 15.55 million person-times of evacuation organised on the back of them.
The lead-time number matters operationally. A flood wave takes time to travel, and a release takes time to reach the next city. Knowing seven days earlier lets engineers draw down storage that would otherwise be spilled at the peak, and lets officials move people in daylight.
What it did in actual floods
The ministry credits the system with supporting the response to the Haihe basin's "23·7" extreme flood in 2023 and the Beijiang (北江) flood of 2024, which it describes as the largest on that river since 1915. It also points to the 2025 Yangtze season, when inflow was below normal: the models helped schedule 58 controlling reservoirs so they ended the flood season holding 108.1 billion cubic metres, which the ministry calls a record.
A caution on the Haihe case. That flood still inundated large areas of Hebei, and reservoir scheduling is one input among many. "Provided support" is the ministry's phrasing, and the right level of claim to repeat.
Where the hard part actually is
Big rivers are the easy case: gauged, modelled, and slow. The hard cases are the ones that kill people — small mountain catchments where a burst of rain turns a dry gully into a wall of water in under an hour. China issued about 1.2 million flash-flood warnings between 2021 and 2025. At that volume, the bottleneck is less detection than whether anyone acts on the message in time. Ten extra days on the Yangtze does nothing for a village in a catchment no radar reaches.
Then there is maintenance. A twin is only as good as its last survey. Channels silt, new roads change runoff, a suburb paves a floodplain — each change quietly degrades a model that was accurate when calibrated.
Honest limitations
- Every figure here comes from the Ministry of Water Resources, stated at its own briefing or written replies. There is no independent audit of the 90% claim, and the ministry does not publish how "accuracy" is defined — per station, per event, or per threshold crossing.
- The jump from three days to ten cannot be attributed to software alone. More radars, more satellites and 136,000 reporting stations feed the same improvement; separating the model's contribution from the sensor build-out is not possible with public data.
- This article covers large-river forecasting. It does not evaluate flash floods, urban drainage or drought, where performance is different.
- No peer-reviewed benchmark of these operational systems is cited here, and the institutional setup — one ministry owning data, models and reservoir dispatch — is specific to China.
What readers can do now
- If you live in a flood-exposed basin, find out which agency issues your warnings and how long its typical lead time is. Ten days is only useful if the message reaches you in a form you act on.
- Ask for uncertainty, not a single number. "Peak between 8.2 and 9.6 metres" tells you what to prepare for; one figure invites false confidence.
- If you work in infrastructure or insurance, treat river models as maintenance items. The question is not whether a basin has a twin, but when its cross-sections were last resurveyed.
