The Zhedong Water Diversion Digital Twin: From Physical Water Network to Intelligent Scheduling

Introduction

In September 2023, General Secretary Xi Jinping, during his inspection of Zhejiang, emphasized the need to "accelerate the construction of the national water network and provide solid water-security support for the comprehensive building of a modern socialist nation." As a pilot zone for provincial water-network construction and one of the first provinces to pilot Digital Twin Water Network construction, the Zhedong Water Diversion Project in Zhejiang Province carries the responsibility of supplying water to 17.5 million people in the core area of the Yangtze River Delta.

This article, drawing on the Zhedong Water Diversion digital twin's construction experience, lays out how to use modern information technology to build a trans-basin, multi-objective, intelligent Digital Twin Water Network — and offers reference for fellow practitioners.

---

I. Project Background: A "Lifeline" Connecting Three Major Basins

1.1 The Physical Water Network at a Glance

The Zhedong Water Diversion is the largest trans-basin water transfer project in Zhejiang Province, with the following standout characteristics:

Spatial scale:

• The diversion main line runs 294 km
• It connects the Qiantang River, Cao'e River, and Yongjiang basins
• It covers 18 counties (cities, districts) across the four cities of Hangzhou, Shaoxing, Ningbo, and Zhoushan

Water transfer scale:

• Average annual diversion volume of 890 million m³ (about 35 West Lakes)
• The Xiaoshan Hub is designed for 50 m³/s
• The North Cao'e Diversion (Sanxing Gate) is designed for 60 m³/s
• The South Cao'e Diversion (Shangyu Hub) is designed for 40 m³/s

Socioeconomic benefits:

• Beneficiary population about 17.5 million (26.7% of Zhejiang Province)
• Beneficiary GDP about CNY 2.14 trillion (36.6% of the province)
• Cumulative diverted water exceeds 5 billion cubic meters

1.2 Diversion Routes and Engineering System

Diversion path (Qiantang River → Xiao-Shao-Yong Plain):

Fuchun River / Qiantang River
  ↓ [Xiaoshan Hub] 50 m³/s
Xiao-Shao Plain river network (Hangyong Canal, Sanjiang River)
  ↓ [New Sanjiang Gate, Mashan Gate]
Cao'e River Tide Gate upstream waterway (regulating storage 23.4 million m³)
  ├→ [Sanxing Gate] 60 m³/s → North Cao'e Line → Yu-Yu-Ci area
  └→ [Shangyu Hub] 40 m³/s → South Cao'e Line → Ning-Zhou area

Core engineering system:

1. Headworks: Xiaoshan Hub (gravity flow + pumped flow)
2. Intermediate hub: Cao'e River Tide Gate (normal storage 146 million m³)
3. Conveyance channels:
   • Xiao-Cao conveyance corridor (62 km)
   • North Cao'e Line (85 km, serving Shangyu, Yuyao, Cixi)
   • South Cao'e Line (93 km, serving Yuyao, Ningbo, Zhoushan)
4. Control gates and stations: 17 major gates/stations + 4 regulating reservoirs

---

II. The Core Problem: Challenges of Traditional Scheduling

2.1 Data Silos and Information Barriers

Status quo:

• Zhejiang Province had already built systems for "Water Conservation Digital Application" and "Water Resources Management Digital Application"
• But the systems were fragmented, data resources were scattered, and a unified data-sharing mechanism was missing
• Cross-agency data exchange involving meteorology, emergency response, and others was difficult

Typical scenario:

Schedulers had to manually pull data from multiple systems and assemble a complete picture of the water situation — slow, labor-intensive, and error-prone.

2.2 Empirical Decisions and the Missing "Four Predictions"

Traditional scheduling:

• Reliance on experience: scheduling decisions hinged on the personal experience of the dispatcher
• Lack of foresight: scientific forecasting and warning instruments were missing
• Lagging response: slow reactions to extreme events such as sudden droughts and floods
• Hard to evaluate: difficulty in quantifying the actual effectiveness of scheduling plans

Gaps in the "Four Predictions":

2.3 Insufficient Capacity for Fine-Grained Scheduling

Pain points:

1. Large differences between river districts: 6 priority river districts (upper/lower Yuyao, middle Mazhu, west/middle/east Cixi) have water levels ranging from 1.53 m to 2.7 m, making integrated, precise scheduling difficult
2. Many intake points: a large number of industrial and agricultural intakes line the route, but there is no fine-grained demand forecasting or water allocation optimization
3. Conflicting objectives: flood control, water supply, and ecological replenishment objectives clash, and there is no scientific tool to weigh them

---

III. The Solution: A Digital Twin Water Network Technology Stack

3.1 Overall Architecture: "1 + 3 + 4 + 1"

Based on the Ministry of Water Resources' "Digital Twin Water Conservancy" framework, the Zhedong Water Diversion digital twin is structured as:

┌─────────────────────────────────────────────┐
│         Application Layer (4 scenarios)      │
│  ┌───────┐ ┌───────┐ ┌───────┐ ┌───────┐  │
│  │Safety │ │Joint  │ │Daily  │ │Emerg. │  │
│  │Monitor│ │Sched. │ │Mgmt.  │ │Resp.  │  │
│  └───────┘ └───────┘ └───────┘ └───────┘  │
├─────────────────────────────────────────────┤
│         Digital Twin Platform (3 stores)     │
│  ┌─────────────────────────────────────┐   │
│  │ Data Base Layer (L1/L2/L3 spatial)  │   │
│  │ + Basic / Monitoring / Business /    │   │
│  │   Shared data                        │   │
│  └─────────────────────────────────────┘   │
│  ┌─────────────────────────────────────┐   │
│  │ Model store (hydrology / water res. │   │
│  │ / hydrodynamics / scheduling)        │   │
│  │ + Intelligent AI recognition models  │   │
│  └─────────────────────────────────────┘   │
│  ┌─────────────────────────────────────┐   │
│  │ Knowledge Base (rules / plans /      │   │
│  │ historical scenarios)                │   │
│  │ + Expert experience base             │   │
│  └─────────────────────────────────────┘   │
├─────────────────────────────────────────────┤
│   Sensing Layer (1 sensor net, 113 sites)   │
│  Water level (44) + Quality (19) +          │
│  Flow (5) + Gate/pump status (23) +         │
│  Safety monitoring (22) = 113               │
└─────────────────────────────────────────────┘

Design philosophy:

• One sensing network: comprehensive coverage of water condition, engineering condition, water quality, and video surveillance
• Three data stores: data base layer, model store, knowledge base
• Four application scenarios: safety monitoring, joint scheduling, daily management, emergency response
• One support system: cybersecurity, standards, organizational mechanisms

3.2 Data Base Layer: The Digital Mirror of the Physical Water Network

3.2.1 Multi-Source Data Fusion

Five data categories aggregated:

3.2.2 A Three-Tier Geospatial Data System

Tiered modeling strategy:

• Tier L1 (macro situational view): provincial DEM, satellite imagery, administrative boundaries

  • Precision: 5 m – 10 m
  • Use: regional situational displays, macro water-volume distribution

• Tier L2 (mid-level management): river network vectors, water conservancy infrastructure, oblique photography

  • Precision: 1 m – 5 m
  • Use: river-network water-level simulation, diversion flow-field analysis

• Tier L3 (micro operations): BIM models for key projects (Xiaoshan Hub, Shangyu Hub, Cao'e River Tide Gate)

  • Precision: 0.01 m – 0.1 m
  • Use: structural safety monitoring, equipment operations and maintenance

3.2.3 Data Governance and Quality Control

Data governance pipeline:

Raw data
  ↓ [validation]
  ├─ Completeness (missing-value detection)
  ├─ Consistency (sanity checks)
  └─ Accuracy (comparison with history)
  ↓ [cleaning]
  ├─ Outlier handling (3σ rule)
  ├─ Deduplication
  └─ Format standardization (units, encoding)
  ↓ [fusion]
  ├─ Spatiotemporal alignment
  ├─ Coordinate system conversion
  └─ Multi-source data linkage
  ↓ [data services]
Standardized data products

Key technologies:

• Real-time stream processing: Kafka-based message queues for second-level data response
• Spatiotemporal database: PostGIS supporting spatial queries and time-series analysis
• Data lineage tracing: recording data origin, processing steps, and quality assessments

3.3 Model Store: The Engine of the Intelligent Brain

3.3.1 Hydrological Forecasting Models

Runoff forecasting (LSTM deep-learning model):

• Inputs: historical runoff, observed/forecast rainfall, evaporation, groundwater level
• Outputs: 3- to 15-day inflow forecasts
• Accuracy: Nash-Sutcliffe Efficiency (NSE) > 0.85

# Pseudocode: LSTM runoff forecasting model
class RunoffForecastModel:
    def __init__(self):
        self.lstm = LSTM(units=128, return_sequences=True)
        self.dense = Dense(units=1)
    
    def predict(self, rainfall, evaporation, groundwater, history_runoff):
        # Feature engineering
        features = self.feature_engineering(
            rainfall, evaporation, groundwater, history_runoff
        )
        # LSTM prediction
        forecast = self.lstm(features)
        forecast = self.dense(forecast)
        return forecast  # Daily runoff for the next 15 days

Water demand forecasting (multivariate time series):

• Industrial water use: based on historical patterns + workday/holiday markers
• Agricultural water use: based on weather forecasts + crop area + irrigation regime
• Domestic water use: based on population + seasonal patterns + temperature forecasts

Technical highlights:

• Integrates weather forecast data to predict demand 7–15 days ahead
• Differentiates 6 river districts to deliver fine-grained zonal forecasts
• Accounts for shifts in demand under extreme events such as droughts and floods

3.3.2 Hydrodynamic Models

One-dimensional river-network hydrodynamic model (Saint-Venant equations):

$$\begin{cases} \frac{\partial A}{\partial t} + \frac{\partial Q}{\partial x} = q \\ \frac{\partial Q}{\partial t} + \frac{\partial}{\partial x}\left(\frac{Q^2}{A}\right) + gA\frac{\partial Z}{\partial x} + \frac{gn^2Q|Q|}{A^2R^{4/3}} = 0 \end{cases}$$

Capabilities:

• Simulates water-level and velocity changes in the river network during diversion
• Simulates response time for water levels under different scheduling plans (6 to 72 hours)
• Computes water-volume allocation under coordinated gate-station scheduling

Model parameters:

• Channel roughness: $n = 0.025$ (Xiao-Shao Plain river network)
• Time step: $\Delta t = 300s$
• Spatial step: $\Delta x = 500m$

Validation accuracy:

• Water-level error: $< 5\,\text{cm}$ (representative stations)
• Flow error: $< 10\%$

3.3.3 Water Resources Scheduling Model

Multi-objective optimization model:

$$\begin{aligned} \max \quad & F = w_1 \cdot F_{\text{supply}} + w_2 \cdot F_{\text{ecology}} + w_3 \cdot F_{\text{economy}} \\ \text{s.t.} \quad & V_{\min} \leq V(t) \leq V_{\max} \quad \forall t \\ & Q_{\min} \leq Q(t) \leq Q_{\max} \quad \forall t \\ & \sum Q_{\text{out}} = \sum Q_{\text{in}} \quad \text{(water balance)} \end{aligned}$$

Constraints:

• River-network water-level constraints (safe level, alert level)
• Intake chloride constraint ($< 250\,\text{mg/L}$)
• Ecological flow constraint (no less than 10% of the long-term mean)
• Gate-station operating constraints (opening range, minimum dwell between switches)

Solution methods:

• Routine scheduling: Dynamic Programming (DP)
• Emergency scheduling: Particle Swarm Optimization (PSO) + constraint handling
• Real-time optimization: Model Predictive Control (MPC)

# Pseudocode: MPC real-time scheduling optimization
class MPCScheduler:
    def __init__(self, prediction_horizon=72, control_horizon=24):
        self.Np = prediction_horizon  # Prediction horizon (hours)
        self.Nc = control_horizon     # Control horizon (hours)
    
    def optimize(self, current_state, demand_forecast, inflow_forecast):
        # Rolling optimization
        while True:
            # 1. Predict the next 72 hours of state
            predicted_states = self.predict(current_state, inflow_forecast)
            
            # 2. Optimize the next 24 hours of control variables (gate openings)
            optimal_control = self.solve_optimization(
                predicted_states, demand_forecast
            )
            
            # 3. Apply the first control action
            self.execute_control(optimal_control[0])
            
            # 4. Update state, roll forward
            current_state = self.update_state()
            time.sleep(3600)  # Wait 1 hour

3.3.4 Intelligent AI Recognition Models

Computer vision applications:

1. Video-based water-level recognition (YOLO object detection + scale conversion)

   • Reads water-staff markings
   • Accuracy: $\pm 1\,\text{cm}$
   • Use: unmanned water-level stations

2. Water hyacinth / blue-green algae recognition (semantic segmentation)

   • Real-time monitoring of channel water quality anomalies
   • Alert threshold: coverage > 30%
   • Use: aquatic ecology safety alerts

3. Gate-opening recognition (keypoint detection)

   • Automatic detection of gate open/close status
   • Verifies execution of dispatch orders

Technical architecture:

Video stream
  ↓ [edge compute device]
  ├─ Real-time inference (TensorRT acceleration)
  ├─ Anomaly detection
  └─ Result upload
  ↓ [cloud platform]
Aggregated analysis + alert publication

3.4 Knowledge Base: Captured Experience and Intelligent Recommendation

3.4.1 Knowledge System Construction

Six knowledge bases:

3.4.2 Knowledge Graph and Intelligent Reasoning

Knowledge graph construction:

[Xiaoshan Hub]
  --|diverts to|--> [Xiao-Shao Plain river network]
  --|flows through|--> [New Sanjiang Gate]
  --|joins|--> [Cao'e River Tide Gate upstream waterway]
  --|chloride-limited at|--> [intake chloride < 250 mg/L]
  --|scheduling basis|--> [Fuchun Power Station release flow]

[Sanxing Gate]
  --|capacity|--> [60 m³/s]
  --|service area|--> [North Yuyao Plain upper district]
  --|downstream gates|--> [Puqian Gate, Zhatouyan Gate, Moushan Gate]
  --|allocation ratio|--> [Shangyu 40%, Yuyao 25%, Cixi 35%]

Intelligent reasoning example:

User query: "The Cao'e River Tide Gate is currently at 3.5 m. Can we open the Sanxing Gate?"

Reasoning engine:
1. Query the business rules:
   Rule 1: "Sanxing opening condition: Cao'e River Tide Gate level > 3.6 m"
   
2. Evaluation: current level 3.5 m < 3.6 m
   
3. Conclusion: opening condition not met
   
4. Recommendation:
   - Continue diverting from Xiaoshan Hub
   - Expected to reach 3.6 m in 2 hours
   - Prepare Sanxing Gate for opening in advance

---

IV. Core Capabilities: Precise Scheduling Driven by the "Four Predictions"

4.1 Water Resources Scheduling — The Four Predictions

4.1.1 Forecast

A multi-timescale forecasting system:

Fine-grained forecasting for 6 river districts:

Taking the upper Yuyao Plain district as an example:

# District-level demand forecast
def forecast_demand(region="Upper Yuyao", horizon=7):
    # 1. Pull historical water-use data
    history = get_history_water_use(region, days=90)
    
    # 2. Pull weather forecast
    weather = get_weather_forecast(region, days=horizon)
    
    # 3. Feature engineering
    features = {
        'date': date,
        'temperature': weather.temp,
        'rainfall': weather.rainfall,
        'workday_flag': is_workday,
        'irrigation_demand': calc_irrigation_demand(weather, crop_type),
        'historical_same_period_mean': history.mean(),
        '7-day_moving_avg': history.rolling(7).mean()
    }
    
    # 4. ML prediction
    demand_forecast = ml_model.predict(features)
    
    return demand_forecast  # Daily demand for the next 7 days

Water resources situation forecasting:

By combining inflow and demand forecasts, predict future river-network storage:

$$V(t+\Delta t) = V(t) + Q_{\text{in}}(t)\Delta t - Q_{\text{out}}(t)\Delta t - Q_{\text{evap}}(t)\Delta t$$

4.1.2 Alert

Tiered alerting system:

Alert indicator system (using upper Yuyao Plain as an example):

upper_yuyao_plain:
  representative_station: Linshanshang
  safe_level: 2.7m
  alert_level: 2.4m
  thresholds:
    red: < 2.5m  # Severe supply shortage
    orange: < 2.6m  # Tight supply
    yellow: < 2.65m # Moderately tight supply
    blue: < 2.7m and forecast continued decline over next 3 days
  responses:
    red: 
      - Puqian Gate at maximum diversion (36 m³/s)
      - Activate emergency water source (reservoir release)
      - Restrict non-domestic water use
    orange:
      - Puqian Gate increased diversion (30 m³/s)
      - Pre-stage emergency water sources
    yellow:
      - Puqian Gate moderate diversion (24 m³/s)
    blue:
      - Notify relevant units to prepare in advance

Intelligent alerting algorithm:

def intelligent_alert(region, forecast_data):
    # 1. Get current state
    current_level = get_water_level(region)
    current_storage = get_water_storage(region)
    
    # 2. Predict future state
    future_level = forecast_data['water_level']
    future_demand = forecast_data['demand']
    future_supply = forecast_data['supply']
    
    # 3. Supply-demand gap analysis
    deficit = future_demand - future_supply
    deficit_ratio = deficit / future_demand
    
    # 4. Composite judgment
    if current_level < 2.5 or deficit_ratio > 0.3:
        return "Red Alert"
    elif current_level < 2.6 or deficit_ratio > 0.2:
        return "Orange Alert"
    elif current_level < 2.65 or deficit_ratio > 0.1:
        return "Yellow Alert"
    elif future_level.min() < 2.7:
        return "Blue Alert"
    else:
        return "Normal"

4.1.3 Simulation

Multi-scenario simulation comparison:

For the same alert scenario, generate 3–5 candidate scheduling plans and run rapid model simulations.

Example scenario: upper Yuyao Plain district at 2.55 m (Orange Alert), forecast continued low rainfall over the next 7 days.

Recommended: Plan B

• ✅ Restores level to safe range
• ✅ No emergency source required
• ✅ Downstream Cixi area also benefits (additional 5 m³/s)

Simulation visualization:

Upper Yuyao Plain water-level simulation (Plan B)

Level (m)
2.80 ┤                            ╭────────
2.75 ┤                      ╭─────╯
2.70 ┤                ╭─────╯
2.65 ┤          ╭─────╯
2.60 ┤    ╭─────╯
2.55 ┼────╯  ← current
2.50 ┤
     └┬─────┬─────┬─────┬─────┬─────┬─────┬
      0     1     2     3     4     5     6 (day)

Puqian Gate diversion flow (m³/s)
40  ┤ ████████████████████████████  ← 36 m³/s
30  ┤
20  ┤
10  ┤
0   └┬─────┬─────┬─────┬─────┬─────┬─────┬
     0     1     2     3     4     5     6 (day)

Key technologies:

• Accelerated computation: GPU parallelization completes 7-day simulation of 3–5 plans in 15 minutes
• Uncertainty analysis: Monte Carlo simulation accounts for uncertainty in rainfall and demand

4.1.4 Plan

Dynamic plan generation:

Based on forecast, alert, and simulation results, automatically generate scheduling plans:

Scheduling Plan ID: ZD-2025-0315-001
Generated: 2025-03-15 10:30
Alert level: Orange Alert
Region: upper Yuyao Plain
Validity: 2025-03-15 ~ 2025-03-22 (7 days)

# Scheduling objectives
Goal 1: Restore upper Yuyao level above 2.7 m
Goal 2: Safeguard domestic water supply in Yuyao
Goal 3: Honor diversion needs for Cixi

# Scheduling actions
Action 1: Puqian Gate diversion
  - Flow: 36 m³/s
  - Duration: 24 hours/day
  - Cumulative: 3.11 million m³/day

Action 2: Sitang Gate, Qitang Gate diversion
  - Sitang flow: 20 m³/s
  - Qitang flow: 20 m³/s
  - Duration: 20 hours/day

Action 3: Intake control
  - Restrict large industrial intakes (prioritize domestic supply)
  - Pause landscape water use

# Schedule
Day 1-3: Full-power diversion, Puqian at 36 m³/s
Day 4-5: After level recovers, Puqian to 30 m³/s
Day 6-7: After level stabilizes, Puqian to 24 m³/s

# Forecast outcomes
Day 7: Upper Yuyao forecast level 2.72 m (safe)
Cumulative diversion: 21.77 million m³
Power cost: ~CNY 150,000

# Emergency fallback
If Day 3 level still < 2.6 m:
  - Activate emergency water source (XX reservoir release)
  - Tighten intake controls

4.2 Safe Operation Monitoring

4.2.1 Structural Safety Monitoring

Xiaoshan Hub safety monitoring system (already deployed):

• Deformation monitoring: 8 GPS automated monitoring points
• Seepage monitoring: 12 piezometers
• Stress and strain: 6 rebar stress meters
• Video surveillance: 16 high-definition camera channels

Safety assessment model:

def safety_assessment(structure="Xiaoshan Hub"):
    # 1. Collect monitoring data
    deformation = get_deformation_data(structure)
    seepage = get_seepage_data(structure)
    stress = get_stress_data(structure)
    
    # 2. Anomaly detection
    anomalies = []
    if deformation.max() > threshold_deformation:
        anomalies.append("Deformation exceeded")
    if seepage.gradient > threshold_seepage:
        anomalies.append("Seepage anomaly")
    if stress.max() > threshold_stress:
        anomalies.append("Stress over limit")
    
    # 3. Composite assessment
    if len(anomalies) == 0:
        return "Safe"
    elif len(anomalies) == 1:
        return "Marginally safe", anomalies
    else:
        return "Risk present", anomalies

4.2.2 Water Supply Safety Monitoring

Real-time monitoring metrics:

Supply-demand balance monitoring:

Upper Yuyao Plain supply-demand status (real-time)

Current: Supply roughly matches demand
River-network storage: 18.5 million m³ (85% full)
Daily supply capacity: 1.25 million m³/day
Actual usage: 1.10 million m³/day
Margin: +150,000 m³/day (13.6%)

7-day forecast:
Day 1-3: Supply > demand (+100,000–150,000 m³/day)
Day 4-7: Balanced (±50,000 m³/day)

Status: 🟢 Normal

4.2.3 Water Quality Safety Monitoring

19 water-quality stations under real-time monitoring:

• Routine indicators: pH, dissolved oxygen (DO), turbidity, conductivity
• Pollution indicators: ammonia nitrogen (NH₃-N), total phosphorus (TP), permanganate index (CODMn)
• Biological indicators: chlorophyll a (algal-bloom alert)

Water-quality alert case:

One day in August 2024, a Shaoxing-section monitoring station detected elevated ammonia nitrogen (1.2 mg/L, exceeding the standard).

Alert response timeline:
1. [10:15] System auto-detected the anomaly and issued a Yellow Alert
2. [10:20] Dispatch center verified and activated the water-quality emergency plan
3. [10:30] Traced pollution to upstream illegal industrial discharge
4. [10:45] Coordinated with environmental authorities; ordered the offender to halt and rectify
5. [11:00] Adjusted scheduling: increased Xiaoshan Hub diversion to dilute the pollutant
6. [14:00] Ammonia nitrogen down to 0.8 mg/L (in compliance)
7. [16:00] Alert lifted, routine scheduling resumed

4.3 Emergency Response

4.3.1 Sudden Water Pollution Events

Emergency response flow:

Pollution dispersion model (advection-diffusion equation):

$$\frac{\partial C}{\partial t} + u\frac{\partial C}{\partial x} = D\frac{\partial^2 C}{\partial x^2} - kC$$

• $C$: pollutant concentration
• $u$: flow velocity
• $D$: diffusion coefficient
• $k$: degradation coefficient

4.3.2 Sudden Engineering Failures

Typical scenario: Sanxing Gate jams and cannot be opened normally.

Emergency scheduling plan:

Event: Sanxing Gate failure
Impact: North Cao'e Line cannot divert normally
Response:
  Plan 1: Increase Shangyu Hub diversion
    - Shangyu flow raised to 50 m³/s (10 m³/s above normal)
    - Replenish Yuyao and Cixi via the Yaojiang trunk
    - Assessment: partial relief, limited effect
  
  Plan 2: Activate emergency water sources
    - Siming Lake reservoir supplies Yuyao
    - Cixi switches to local reservoirs
    - Assessment: covers 7–10 days

  Repair plan:
    - Estimated repair time: 48 hours
    - Emergency scheduling lasts at least 2 days

4.3.3 Localized Heavy-Rain Flood Scheduling

Scenario: a typhoon brings heavy rain — 200 mm in 24 hours over Yuyao.

Flood scheduling response:

Flood Dispatch Order No. ZD-FL-2025-0815-001

Forecast: Typhoon "XX" impacts; Yuyao area expects 200 mm in 24 h
Current state: 
  - Upper Yuyao level: 2.85 m (above alert level of 2.7 m)
  - Yaojiang level: 1.45 m (close to alert level of 1.5 m)

Actions:
1. Immediately close all diversion gates
   - Puqian Gate: closed
   - Zhatouyan Gate, Moushan Gate: closed
   - Sitang Gate, Qitang Gate: closed

2. Activate drainage gates
   - Longshanpu drainage station: 6 units (60 m³/s)
   - Linhaipu New Gate: fully open
   - Taojialu New Gate: fully open

3. Lower upstream water level
   - Sanxing Gate: closed, reducing diversion to Yuyao
   - Shangyu Hub: closed, reducing diversion to Yaojiang

4. Real-time monitoring
   - Densify water-level sampling (every 10 minutes)
   - Activate 24-hour duty roster

Goal: Lower upper Yuyao level below 2.7 m

---

V. Innovation Highlights

5.1 Six-District Fine-Grained Scheduling

Innovation: the first domestic implementation of fine-grained joint scheduling across multiple plain river-network districts (6 districts).

Technical challenges:

• Six river districts with sizeable water-level differences (1.53 m to 2.7 m) and complex hydraulic connectivity
• Strong coupling among 17 control gates and stations
• Simultaneous satisfaction of flood-control, supply, and ecological objectives

Solutions:

1. Tiered, district-aware scheduling strategy:

   • Provincial (Qiantang River → Cao'e River): controls overall diversion volume
   • Municipal (Cao'e River → Yuyao/Cixi): controls regional allocation
   • County (within Yuyao/Cixi): fine-grained allocation

2. Coordinated gate-station optimization:

   Objective: minimize supply-demand gaps across districts
   Decision variables: openings and switching schedules of 17 gates
   Constraints: water-level safety, flow limits, ecological flow
   

3. Real-time feedback correction:

   • Recalibrate the forecasting model hourly using observations
   • Dynamically adjust scheduling plans

5.2 Model Platformization and Knowledge Graph

Innovation: built a water-conservancy model platform + knowledge graph for model reuse and intelligent recommendation.

Model registration and sharing mechanism:

Model Marketplace
├─ Hydrological models (12)
│  ├─ Runoff forecasting v2.1 ⭐⭐⭐⭐⭐ (called 586 times)
│  ├─ Flood forecasting v1.8 ⭐⭐⭐⭐
│  └─ ...
├─ Water resources models (8)
│  ├─ Demand forecasting v3.0 ⭐⭐⭐⭐⭐ (called 412 times)
│  ├─ Water balance v2.5 ⭐⭐⭐⭐
│  └─ ...
├─ Hydrodynamic models (5)
└─ Intelligent AI models (6)

Knowledge graph applications:

Scenario: a dispatcher asks, "How should we respond to persistent drought in Yuyao?"

Knowledge graph reasoning:
1. Recognize entities: [Yuyao area, persistent drought]
2. Linked queries:
   - Historical scenarios: 2022 Yuyao drought (similarity 92%)
   - Expert experience: Engineer Wang's principle: "increase Puqian diversion first"
   - Scheduling plans: Plan #156 (effective)
3. Recommendation:
   - Reference Plan #156 from 2022
   - Key actions: Puqian at 36 m³/s + emergency source
   - Expected outcome: drought relieved within 7 days

5.3 Video AI Recognition

Innovation: the first large-scale application of video AI recognition in the domestic water conservancy industry.

Use cases:

1. Video water-level recognition (replacing manual readings)

   • Accuracy: 98.5%
   • Labor saved: 15 stations × 4 readings/day × 5 minutes = 300 person-minutes/day

2. Algae / water-hyacinth recognition (ecological alerts)

   • Real-time monitoring of channel surfaces
   • Auto-alerts on anomalies
   • Three algal-bloom risks were proactively flagged in 2024

3. Gate-opening recognition (scheduling oversight)

   • Auto-verifies dispatch order execution
   • Flags non-compliance automatically

5.4 Physical Dispatch Center

Innovation: built a provincial-municipal-county tiered intelligent dispatch and command center.

Functional layout:

┌─────────────────────────────────────────────┐
│      Intelligent Dispatch Center (physical)  │
│                                             │
│  ┌────────────────────────────────────┐    │
│  │   Big-screen wall (5m × 3m LED)     │    │
│  │  ┌──────┐ ┌──────┐ ┌──────┐        │    │
│  │  │Status│ │Forec.│ │Sched.│        │    │
│  │  └──────┘ └──────┘ └──────┘        │    │
│  └────────────────────────────────────┘    │
│                                             │
│  ┌─────┐  ┌─────┐  ┌─────┐  ┌─────┐       │
│  │Disp.│  │Cond.│  │Forec│  │Tech.│       │
│  │ desk│  │ desk│  │ desk│  │ desk│       │
│  └─────┘  └─────┘  └─────┘  └─────┘       │
│                                             │
│  ┌──────────────────────────────────────┐  │
│  │  Video conference (city/county tie-in)│ │
│  └──────────────────────────────────────┘  │
│                                             │
└─────────────────────────────────────────────┘

Technical highlights:

• Remote centralized control: remotely operates the 17 main gates and stations along the route
• Emergency command: video conferencing, one-click scheduling
• 24/7 duty: continuous monitoring of the water situation

---

VI. Outcomes and Impact

6.1 Substantially Improved Scheduling Efficiency

Comparative data (2023 vs 2025):

Case study: 2024 summer drought response

Background: from July to August 2024, eastern Zhejiang experienced sustained heat and low rainfall, and the Yuyao river-network water level fell continuously.

Traditional mode (a similar event in 2023):

• Detect drought → manual analysis → consultative judgment → produce plan: takes 3 days
• Water level dropped from 2.55 m to 2.45 m, triggering a Red Alert
• Activated emergency water source — additional cost about CNY 500,000

Digital twin mode (2024 actual):

• The system predicted the drought 7 days ahead and auto-issued a Blue Alert
• Generated a scheduling plan that the dispatcher confirmed and executed
• Diversion was increased early; level held at 2.65–2.70 m
• No Red Alert triggered, no emergency source used, CNY 500,000 saved

6.2 Improved Water-Use Efficiency

Outcomes from fine-grained scheduling:

• Diversion plan accuracy: improved from 75% to 92%
• Spillage rate: reduced from 8% to 3% (less ineffective drainage)
• Supply-demand match: improved from 80% to 95%

Economic estimate:

Annual water savings: 20 million m³
At industrial water price CNY 2.5/m³: CNY 50 million/year
Less system construction and operating cost: CNY 10 million/year
Net benefit: CNY 40 million/year
Payback period: ~2.5 years

6.3 Improved Ecological Replenishment

River-lake ecological flow guarantee rate:

• 2023 (traditional mode): trunk-channel ecological flow guarantee rate 68%
• 2025 (digital twin): trunk-channel ecological flow guarantee rate 89%

Water quality improvements:

Algal bloom prevention outcomes:

• In 2024, video AI flagged 3 algal-bloom risks early
• Increased diversion in time prevented major outbreaks
• Reduced water-quality emergency cost by about CNY 2 million/year

6.4 Recognition

• 2024: selected as a best-practice case by the Ministry of Water Resources for Digital Twin Water Network construction
• 2024: awarded Second Prize of the Zhejiang Provincial Science and Technology Progress Award
• 2025: selected as a pilot best-practice by the Ministry of Water Resources for nationwide promotion

---

VII. Lessons and Outlook

7.1 Key Lessons

7.1.1 Top-Down Design Is the Foundation

Lesson: early stovepipe builds led to data silos.

Takeaways:

• ✅ Unify data standards (per the MWR's "Technical Guidelines for Digital Twin Water Conservancy")
• ✅ Unify the platform architecture ("data base layer + model platform + knowledge platform")
• ✅ Unify the API contract (RESTful + WebSocket)

7.1.2 Models Are the Core

Lesson: without professional models early on, the "Four Predictions" were just paperwork.

Takeaways:

• ✅ Develop core models in-house (hydrological forecasting, water resources scheduling)
• ✅ Continually calibrate and validate (use historical data to refine parameters)
• ✅ Platformize models (for sharing, reuse, and rapid iteration)

7.1.3 Data Quality Is the Key

Lesson: early monitoring data had gaps and anomalies that hurt model accuracy.

Takeaways:

• ✅ Improve the sensing system (filled in key sites; 113 monitoring points total)
• ✅ Apply data quality controls (outlier detection, cleansing, missing-value imputation)
• ✅ Make data governance routine (with a data steward role)

7.1.4 Business Integration Is the Goal

Lesson: dispatchers know best whether the system is genuinely usable.

Takeaways:

• ✅ Involve operations staff throughout requirements gathering and testing
• ✅ Design UIs around dispatcher workflows
• ✅ Provide online training and operating manuals
• ✅ Establish a fast-response mechanism for needs (a 2-week iteration cadence)

7.2 Technical Challenges and Solutions

7.2.1 River-Network Hydrodynamic Simulation Accuracy

Challenge: plain river networks have flat terrain, slow flows, and pronounced backwater — making simulation difficult.

Solutions:

1. Refined terrain data: combined multi-beam bathymetry with UAV photogrammetry to build a high-precision river-network terrain
2. Roughness calibration: inverted roughness distribution from historical flood events
3. Boundary condition refinement: incorporated complex boundaries such as gate switching, tidal backwater, and tributary inflow

Result: water-level simulation accuracy improved from ±15 cm to ±5 cm.

7.2.2 Demand Forecasting Uncertainty

Challenge: water-use behavior is influenced by economic, meteorological, and policy factors, and is hard to forecast precisely.

Solutions:

1. Ensemble forecasting:

   • Statistical models (ARIMA): capture historical patterns
   • Machine learning (XGBoost): mine non-linear relationships
   • Deep learning (LSTM): learn long-range dependencies
   • Weighted ensemble: combine strengths

2. Multi-scenario forecasting:

   • Optimistic scenario (demand −10%)
   • Baseline (normal demand)
   • Pessimistic scenario (demand +10%)

Result: demand forecast error reduced from 25% to 15%.

7.2.3 Real-Time Computation Throughput

Challenge: coupled hydrodynamic models across 6 river districts produce a heavy computational load.

Solutions:

1. GPU parallelization: CUDA-accelerated PDE solvers
2. Model order reduction: POD (Proper Orthogonal Decomposition) cut model dimensions from 10,000+ to 500
3. Distributed computation: multi-node parallelism orchestrated with Kubernetes

Result: 7-day simulation time cut from 2 hours to 15 minutes.

7.3 Looking Ahead

7.3.1 Province-Wide Digital Twin (2026–2027)

Goal: extend the digital twin scope from the Zhedong Water Diversion main channel to the entire province.

• Cover the Qiantang, Cao'e, and Yongjiang basins in full
• Cover every county and city in eastern Zhejiang
• Achieve basin–water-network–project three-tier coordination

7.3.2 AI Large-Model Applications (2027–2028)

Directions:

1. Water conservancy LLMs:

   User: "Is now a good time to open the Sanxing Gate?"
   AI: "Based on real-time data, the Cao'e River Tide Gate is at 3.72 m,
        meeting the opening condition (>3.6 m). I recommend opening
        Sanxing with an initial flow of 40 m³/s. Watch the upper
        North Yuyao Plain (currently 2.85 m); avoid exceeding 3.1 m."
   

2. Reinforcement learning for scheduling:

   • Use AlphaGo-style techniques: self-play to learn the optimal scheduling policy
   • Surpass human experts in complex, fast-changing water conditions

3. Digital twin + metaverse:

   • VR for immersive inspections
   • A digital-human dispatcher available 24/7

7.3.3 Cross-Region Coordination (2028+)

Vision: realize integrated water-network scheduling across the Yangtze River Delta.

• Zhedong Water Diversion ↔ Taipu River ↔ South-to-North Water Transfer (East Route)
• Optimization of inter-provincial water exchange
• Joint allocation of basin-wide water resources

---

VIII. Closing Thoughts

The Zhedong Water Diversion digital twin is a vivid practice of the "Digital China" strategy in the water conservancy industry, and a successful application of the "Four Predictions" framework on a trans-basin water network. Through the deep fusion of physical and digital water networks, we have achieved:

✅ A leap from empirical scheduling to intelligent scheduling
✅ A shift from extensive management to fine-grained management
✅ An upgrade from passive response to proactive prediction

That said, building Digital Twin Water Networks is an ongoing journey. Faced with climate change and the rising frequency of extreme events, we will continue to explore and innovate — contributing water-conservancy intelligence to regional water security and high-quality socioeconomic development.

---

References

1. Ministry of Water Resources. Top-Level Design for Smart Water Conservancy Construction[R]. 2021.
2. Ministry of Water Resources. Technical Guidelines for Digital Twin Basin Construction (Trial)[S]. 2022.
3. Department of Water Resources of Zhejiang Province. Implementation Plan for Zhejiang Water Network Construction (2023–2027)[R]. 2022.
4. Department of Water Resources of Zhejiang Province. Implementation Plan for the Digital Twin Zhejiang Water Network[R]. 2023.
5. Li Guoying. Vigorously Advancing Digital Twin Water Conservancy Construction to Provide Strong Support for High-Quality Development of Water Conservancy in the New Era[J]. China Water Resources, 2023(11).

---

About the Author

Tianli Zeng, Master of Water Conservancy Engineering, engineer at the Zhejiang Provincial Institute of Water Conservancy and Hydroelectric Power Planning and Design. His work focuses on water resources optimal allocation, digital twin water conservancy, and smart water networks. He has participated in projects including the Zhedong Water Diversion digital twin and Huzhou's water resources optimal allocation, among other key provincial and municipal initiatives.

📧 Contact: [email protected]
🔗 Personal site: https://tianlizeng.cloud

---

Originality statement: This article is based on real project experience, with some data anonymized. Please do not republish for commercial use without permission.

Last updated: 2025-10-11