Everyone who runs an urban water supply system dreads two calls: one in the middle of the night reporting no water, and one from the water quality team reporting end-of-line turbidity above the limit.
The DMA-A inlet flow meter reading rose from 380 m³/h to 580 m³/h, up 52%. The number sits on the dispatch screen, and the dispatcher’s first reaction is “peak demand has arrived”. But at the same moment, DMA-A inlet pressure fell from 0.32 MPa to 0.22 MPa, down 31%. Flow up and pressure down is the classic signature pair for network leakage, yet the SCADA system only sets threshold alarms on single variables: one triggers when flow exceeds 460, another when pressure falls below 0.28. The two alarms belong to different subsystems, and nobody combines them into a “flow↑+pressure↓” leakage rule. Only after 240 households call in low-pressure complaints does the repair crew leave.
The other problem is trickier: treated water residual chlorine at the treatment plant quietly slid from 0.50 mg/L to 0.08 mg/L. The chlorinator plunger had worn out, but nighttime demand was low, so end-of-line water quality had not deteriorated yet and the dispatch center noticed nothing. Five hours later the low-chlorine water reached DMA-A, combined with backflow from a user-installed booster pump, end-of-line turbidity spiked to 3.2 NTU, an emergency water quality response plan was activated, and 3,200 households lost water for five hours.
These two problems are, in essence, the same thing: the equipment readings are available, but the physical cause-and-effect chain is not connected.
Why anomalies in the water supply network are so hard to trace
Urban water supply is different from discrete manufacturing. In discrete manufacturing, deviations in each process step are mostly absorbed within that step and do not propagate downstream. But a water distribution network is a continuous physical system where pressure, flow, and water quality are tightly coupled, and an anomaly at one node can propagate through the network and amplify.
Specifically, there are three characteristics:
- First, daily demand swings widely, so the normal baseline keeps moving. Morning peak flow from 6-9 is 14,000 m³/h; the late-night trough from 0-5 is only 7,000 m³/h, a 2:1 swing. A 20% rise in flow may simply be the normal morning peak, or it may be leakage. The judgment dispatchers find hardest is whether a rise in flow means growing demand or a broken pipe.
- Second, pressure and flow are inversely coupled. During peak demand, flow in the network is high, friction losses are high, and end-of-line pressure is actually lower. This means low pressure does not necessarily equal trouble; it may just be heavy usage. Pressure and flow must be viewed together to tell normal operation from an anomaly.
- Third, water quality propagation carries an hour-scale lag. Water leaving the plant takes 4-6 hours to travel through the pipe network to the DMA end of line. If treated water residual chlorine falls, the end of line will not reflect it immediately, and by the time end-of-line turbidity spikes and someone responds, the contamination has already happened. This time gap is the window for catching a water quality incident before it reaches customers.
For this kind of continuous system with coupled variables, a moving baseline, and travel-time lag, traditional single-point threshold alarms cannot cover the risk. One alarm for flow over the limit, one for low pressure, one for high turbidity: the three alarms trigger and are handled separately, and nobody sees that they are three symptoms of the same leak.
Three typical problems, three analysis methods
Problem 1: DMA flow is up. Peak demand or a leak? Flow-pressure coupled analysis
The core judgment a dispatcher faces every day: DMA inlet flow is rising. Should crews be sent out?
The traditional approach sets a fixed threshold on flow, for example alarming above 460 m³/h. But that threshold triggers every day during the summer peak and almost never in winter. A more reliable way is to look at the relationship between flow and pressure: when normal usage rises, flow goes up and pressure dips slightly but not by much (higher friction loss); when there is a leak, flow goes up and pressure falls sharply (water is escaping through the break).
TDengine’s process analysis workspace can overlay DMA inlet flow and inlet pressure on the same time axis, so the coupling pattern between the two is visible at a glance. Going further, plot flow against pressure in a scatter chart: under normal operation the points fall in a narrow band, the normal coupling band where high flow pairs with slightly lower pressure. Once a leak appears, the points jump out of this band and move toward the quadrant of high flow and very low pressure, and the separation from normal operation is clear at a glance.
Problem 2: Which DMA zone is leaking? Cross-zone water balance accounting
The basic principle of leakage identification is simple: total water in should equal the total water out across the DMA zones plus reasonable losses. If a zone’s inlet flow is clearly greater than its outlet flow (metered on the customer side) plus reasonable losses, that DMA has a leak.
But in practice, data from the plant outlet flow meter (MFM-01) and the DMA inlet flow meters (DMA-A-FM, DMA-B-FM) sits in different monitoring stations, the timestamps do not line up, and the sampling rates differ, so the calculated difference flips between positive and negative and cannot support any judgment.
TDengine’s asset modeling places the plant outlet flow meter, the booster pump station, and the DMA flow meters under one asset tree with a unified time base. In the process analysis workspace, the distribution imbalance can be computed in real time as MFM-01 flow minus DMA-A-FM flow minus DMA-B-FM flow minus booster-pump-station flow. If that difference holds above 5%, it is a leak signal; and because the zone with the largest deviation is immediately apparent, the leaking DMA zone is located directly, with no need to inspect section by section.
Problem 3: Treated water quality is off. When will it reach the end of line? Delay-propagation chain tracing
In a water quality incident, the bigger operating problem is often not detecting the anomaly at the plant outlet, but judging when and how seriously the end of line may deteriorate after it is detected. Treated water residual chlorine is 0.08 mg/L, 0.12 below the 0.20 mg/L safety floor, but water takes five hours to reach DMA-A, and there is network decay along the way (end-of-line residual chlorine ≈ plant outlet residual chlorine × 0.85), so the dispatch center must complete its judgment and response within that five-hour window.
The analysis path TDengine provides is: plant outlet trend warning → delay estimation to locate the transmission window → end-of-line vs. plant outlet comparison to confirm the contamination source.
When the sliding mean of residual chlorine at the plant outlet water quality analyzer (MQM-01) begins to drift, the AI anomaly detection algorithm can recognize the directional shift before residual chlorine touches the hard threshold, 1-2 hours ahead of a threshold alarm. With the warning in hand, the same-period trend from the DMA-A water quality analyzer (DMA-A-QM) shows end-of-line residual chlorine falling about five hours behind the plant outlet: that is the network transmission delay. Comparing plant outlet turbidity (still 0.3 NTU, normal) with end-of-line turbidity (spiked to 3.2 NTU) confirms that the contamination source is not at the treatment plant but inside DMA-A. The treated water leaving the plant is clean and the end of line is dirty, which means something in between went wrong (a user-installed booster pump caused backflow contamination).
The analysis loop in a typical anomaly scenario
Scenario 1: Branch-pipe leakage in DMA-A: flow-pressure coupling → water quality confirmation → pump response quantification → end-of-line service assessment
At 11:00 on day 4, a crack opened at the DMA-A branch-pipe joint under the combined effects of long-term corrosion and water hammer. The leak did not happen all at once; it accumulated gradually: first flow ticked up and pressure ticked down, a few hours later flow spiked and pressure crashed, and finally the clear-water reservoir level dropped sharply and end-of-line customers lost water.
Step 1: Detection. DMA-A inlet flow deviates from the historical pattern for the same period.
The hourly average flow at DMA-A-FM kept rising from 380 m³/h and reached 420 m³/h at 11:55 (+10%). Looking at flow alone, a 10% deviation during the morning peak could entirely be explained by normal demand growth. But the AI anomaly detection algorithm does not compare against a fixed threshold; it compares against the historical pattern for the same hourly period. This is the morning flat period, where the historical mean is 380 m³/h with a standard deviation of 25 m³/h, so the current value is already 1.6 sigma off and the trend is still climbing.
Step 2: Cross-check. Pressure and flow form the “flow↑+pressure↓” leakage signature pair.
The same-period trend from DMA-A-PM shows pressure falling from 0.32 MPa to 0.30 MPa (-6%). Flow up 10% and pressure down 6%, opposite directions and reasonable magnitudes: this is a typical leak signature, not peak demand (during peak, the pressure drop is usually no more than 3%). If usage had increased, pressure would not fall this much.
Step 3: Root cause. The simultaneous rise in turbidity confirms the leak is inside DMA-A.
The turbidity trend from DMA-A-QM shows turbidity rising from 0.5 NTU to 0.9 NTU (+80%). This is the decisive criterion: if the cause were simply more water coming in from upstream (peak demand), turbidity would not rise. A rise in turbidity means the leak point is drawing in outside particulates: the negative pressure inside the pipe at the crack pulls soil particles from outside the pipe into the flow. At this point, the leak is confirmed to be inside DMA-A, not an upstream problem.
Step 4: Quantify the impact. Clear-water reservoir level and pump current show the severity of the leak.
As the leak kept growing, the delivery pumps compensated for the pressure drop: the DP-02 motor current rose from 95 A to 132 A (+39%) and the standby pump DP-03 was brought in by interlock. The clear-water reservoir CW-01 level fell from 4.2 m to 3.0 m (-29%), a rate far above normal. From the reservoir level drop rate and the pump current deviation, the leak flow can be estimated: DMA-A inlet flow peaked at 580 m³/h against a normal 380 m³/h, so the leak flow is about 200 m³/h. End-of-line pressure EPS-01 fell from 0.18 MPa to 0.12 MPa and already affects water supply to high-rise users.
Scenario 2: Treated water quality anomaly with end-of-line backflow contamination: trend warning → delay tracing → source location → emergency response
At 00:15 on the morning of day 6, the chlorinator plunger at the treatment plant began to wear, and treated water residual chlorine dipped from 0.50 mg/L to 0.42 mg/L (-16%). At this point plant outlet pressure and flow were entirely normal; only the water quality indicator showed a directional shift. Nighttime demand was low, so end-of-line water quality was not yet affected and the dispatch center had no alarms.
Step 1: Early warning. The AI anomaly detection algorithm recognizes the trend shift in treated water residual chlorine.
The residual chlorine sliding mean at MQM-01 fell 16% during the H120 period. This is not a random fluctuation; it is a continuous downward trend. The AI anomaly detection algorithm recognizes this directional shift and issues a trend warning before residual chlorine touches the 0.40 mg/L threshold alarm line, about 1-2 hours ahead of a traditional threshold alarm.
Step 2: Propagation. The low-chlorine water reaches the DMA-A end of line five hours later.
During the H125 period (05:30), residual chlorine at DMA-A-QM began to fall, from 0.42 to 0.32 mg/L. Overlaying the residual chlorine trends from MQM-01 and DMA-A-QM shows the end of line lagging the plant outlet by about five hours, forming the plant-outlet-to-end-of-line water quality propagation chain. The delay matches the pipe length and flow velocity, confirming it is the propagation of the same event rather than two independent anomalies.
Step 3: Contamination source location. Plant outlet turbidity is normal while end-of-line turbidity spikes, so the contamination is inside DMA-A.
Plant outlet turbidity at MQM-01 held at 0.3 NTU (normal), but end-of-line turbidity at DMA-A-QM spiked to 3.2 NTU. If the treated water itself were dirty, plant outlet turbidity would rise first. It did not rise, and end-of-line turbidity did, which means the contamination source is not at the treatment plant. Combined with the transient negative-pressure record at DMA-A-PM (0.05 MPa), the cause is a user-installed booster pump that created backflow during a low-pressure period and drew outside water into the network.
Step 4: Emergency response. Switch to the standby chlorinator + isolate the affected zone + flush the network.
At 10:00 the dispatch center activated the emergency water quality response plan. The treatment plant switched to the standby chlorinator and residual chlorine at MQM-01 began to recover. DMA-A was isolated urgently (flow dropped from 380 to 80 m³/h) and network flushing began. At 15:00 all water quality indicators returned to normal and supply was gradually restored. From the plant outlet trend warning to the end-of-line contamination outbreak there was a five-hour window. If the AI warning in step 1 had triggered earlier and dispatch had responded sooner, that window would have been enough for preventive action: switching the chlorinator, notifying hospitals and other key users to store water, and avoiding the emergency shutdown.
Water supply network monitoring needs causal visibility through pressure, flow, and water quality, not more thresholds
Digitizing an urban water supply network is not mainly about how many sensors are connected or how many alarm thresholds are set. It depends on whether the data can reconstruct complete causal chains, from leak → pressure drop → water quality change → pump compensation → level drop → end-of-line water outage, and from chlorination anomaly → low treated water chlorine → delay propagation → end-of-line contamination → backflow outbreak → emergency shutdown.
- Flow-pressure coupled analysis: from alarming only when flow exceeds a limit, to flow-pressure coupling pattern recognition plus scatter-chart coupling band deviation, distinguishing peak demand from a pipe leak.
- Cross-zone water balance accounting: from data across monitoring stations not lining up, to real-time computation of the distribution imbalance on a unified time base, locating the leaking DMA zone directly.
- Delay-propagation chain tracing: from discovering the problem only when end-of-line turbidity exceeds the limit, to plant outlet trend warning plus a five-hour transmission window plus plant outlet vs. end-of-line comparison to locate the contamination source, turning passive response into early interception.
TDengine organizes pressure, flow, and water quality data scattered across the treatment plant, booster pump stations, and the distribution network into a unified network asset system through Industrial Ontology modeling, linking single-point alarms into explainable process judgments, so dispatchers can answer three questions the moment flow starts to deviate: is it a leak or peak demand? Which zone? How severe?
TDengine comes with a high-performance, distributed time-series database, Industrial Ontology modeling and an Industrial Agent Runtime, providing a full-stack solution for industrial data streams from collection and storage to real-time analytics, visualization, event management, and root-cause analysis. To learn more about TDengine, visit https://www.tdengine.com/ and download it for free.
Try it yourself
Install and deploy TDengine Visit the TDengine Download Center, select TDengine All-in-One, choose the deployment platform and architecture that matches your environment, and follow the guided steps to complete the installation.
Load the sample data
On first activation, on the sample data loading screen, select Urban Water Supply Network Monitoring and wait for loading to complete.
If you have already activated the product, click your avatar in the top-right corner, select the Management Console, choose Sample Data on the left, then select Urban Water Supply Network Monitoring to load it. Wait a few minutes for loading to finish, and you are ready to explore.
Urban Water Supply Network Monitoring
