Balancing flows in a hydronic network, from domestic heating to data centre cooling

1. The problem

1.1 One source, many consumers

Regardless of the application, every hydronic network is the same object: two collectors, a source that circulates the fluid between them, and a set of parallel branches that each take a share of the flow. A branch can be a radiator, a fan coil, an apartment riser, or a server bay. The requirement always boils down to: each branch must receive its design flow and maintain it regardless of the state of the other branches.

Two distinct failures lie behind this requirement. The static imbalance appears at the design point: when everything is working, the branch closest to the source receives more pressure than the furthest and therefore takes more flow. This is a distribution problem, resolved once during commissioning. The dynamic coupling appears during a change in configuration: when a branch closes or a pump switches, the total flow and pressure losses in the network change, and all the other branches diverge. It is a stability issue that determines the choice of the branch balancing device.

Only the permissible tolerance and the consequence of its exceedance change, depending on the application. In a dwelling, an underfed radiator results in a cold room. In a DLC network, insufficient flow causes throttling of a GPU: the guidelines aim for ± 5% on the branch flow, typically 1.5 L/min per kW for a temperature rise of about 10 °C (D2).

1.2 The physics of flow sharing

The entire argument is contained in one figure and two equations.

Branch law. A branch equipped with a fixed adjustment device behaves like a hydraulic resistance. Its flow is

ṁ_j = K_v √(Δp_j)     with     Δp_j = Δp_set − Δp_dist

where K_v encompasses everything contained in the branch (emitter, hoses, fittings, and valve), and Δp_dist is the pressure loss of distribution between the measurement point of the regulation and the branch take-off.

The essential point lies in the structure of this expression. The flow of a branch depends on only two quantities: its own resistance, invariant for a given material, and the Δp available at its terminals. If the network guarantees a Δp constant at the terminals of the branch, its flow is automatically constant, without any intelligent device, and the state of the other branches then has no influence. The independence from the pressure that a dedicated valve ensures locally can therefore be ensured globally by the network itself.

The residual coupling comes from the fact that Δp_dist is non-zero and depends on the total flow, thus on the state of the other branches. By differentiating the branch law at K_v constant:

δṁ_j / ṁ_j = − ½ · δ(Δp_dist) / Δp_j

A variation in the pressure loss of distribution affects the branch flow by being divided by two and related to the Δp of the branch. The worst case goes from full load to a load close to zero, where Δp_dist disappears entirely. By setting

α = Δp_dist^max / Δp_branch

the maximum flow drift of a branch, between the full network and the nearly empty network, is ε_max ≈ α / 2.

This quantity α is the exact analogue of the authority of a valve in classic regulation. It expresses the fraction of the total pressure carried by the variable part of the circuit. A high authority characterises a network where the branch dominates and where the distribution behaves almost like an ideal connection, a regime in which balancing no longer poses a difficulty.

2. Three levers, not one

The above criterion directly indicates where to act: reduce α, stabilise Δp_set, or make the branch insensitive to the pressure applied to it.

These levers substitute for one another, which is why the field presents such a diversity of solutions. A network with generous collectors and regulated pumps can make do with manual valves; a network with small diameter risers and fixed speed pumps will need to use pressure-independent valves on each branch. The right question is never "do we need a PICV?", but "what is the value of α, and Δp_set is it maintained?".

2.1 Lever 1: the network

Oversizing the distribution is the most direct and robust means, as it has no mode of failure. Its cost is based on steel and fluid volume, not on operation. The two usual rules are a speed limit, less than 1.0 m/s in a DLC collector according to D4 and less than 1.5 m/s in a bay collector according to D5, and the pressure drop ratio α ≤ 1/40. OCP notes in the same spirit that a DN 50 branch carries four times the flow of a DN 25 at the same speed, and that the pressure drop varies with the square of the speed, hence its recommendation to prefer large diameters for "significantly reduce head-loss and balancing issues" (D2), that is to say, to significantly reduce pressure losses and balancing difficulties.

The reverse return, or Tichelmann loop, equalises the total developed length (outgoing plus return) of each branch: the first branch on the outgoing is the last on the return. The static imbalance disappears by design, without adjustment. It is the reference solution when it is feasible; OVHcloud notes that it is not always the case in large installations, « as it requires space and support for a long and large pipe » (translation: because it requires space and supports for a long and large diameter pipe), and it is precisely here that a branch control valve is used (D3). Note that the reverse return addresses the static imbalance, but not the dynamic coupling: it equalises the paths without making them insensitive to the total flow.

The ring topology, supplied at two points, acts on both fronts: it halves the flow carried by each section, thus the distribution pressure loss, and reduces the gap between the best and worst positioned take-offs. It constitutes a common architecture for large secondary loops.

2.2 Lever 2: pump regulation

This is the lever that validates the argument of § 1.2, and it deserves to be explained as it is too often taken for granted.

A fixed-speed pump does exactly the opposite of what is desired. Its characteristic curve decreases: when the total flow decreases, the manometric head increases. The closure of a branch therefore applies more pressure to the others, which then draw more flow than expected. The figure below quantifies the effect on a four-branch network.

This textbook case serves to justify independent pressure limiters and is regularly presented as general proof. It is not: it assumes a fixed-speed pump. With a pump regulated on Δp, the residual drift is no longer that of the pump curve but only that of the distribution pressure loss, namely the term α / 2 of § 1.2, an order of magnitude lower.

The measurement point of the Δp is the only discrete choice that determines the quality of the result. Three strategies coexist:

• Measurement at the source. The simplest solution, adopted by most prefabricated units. The Δp is maintained upstream of the collectors, so the branches experience the full variation of Δp_dist. This is the case covered by the criterion of § 1.2.
• Remote measurement at the disadvantaged point. The sensor measures the pressure at the last take-off. The term Δp_dist cancels out for the furthest branch and changes sign for the others: the regulation then overestimates the pressure instead of underestimating it. The coupling becomes negligible, at the cost of a wired remote sensor.
• Sliding setpoint, where Δp_set is recalculated from the total measured flow in order to analytically compensate for the distribution pressure loss. No remote sensor, but a hydraulic model embedded in the regulation.

2.3 Lever 3: the branch balancing device

The same valve body can belong to two different families depending on whether it has an actuator or not, which explains a large part of the terminological confusion. The table below sets the terms.

Terminology. The two central families, PIBV and PICV, correspond physically to the same product; only the actuator distinguishes them. Caleffi indicates this unambiguously for its valve 145: « With the green manual adjustment knob as shown, this valve is an adjustable PIBV […] Adding an actuator changes the valve from an adjustable pressure-independent balancing device, to a pressure-independent control valve (PICV) » (R1, p. 25 and 26). Translation: with the green manual adjustment knob, this valve is an adjustable PIBV; adding an actuator transforms it into a pressure-independent control valve, that is to say a PICV. Common synonyms for the PIBV: automatic balancing valve, dynamic balancing valve. However, be careful with the acronym PIBCV (Pressure Independent Balancing and Control Valve) : despite the « B », it refers to the motorised, synonymous with PICV at IMI, FlowCon and Danfoss (R3). On the DLC side, OCP uses a simpler nomenclature: manual valves, automatic valves, control valves and (intelligent) valves compatible with telemetry (D2).

2.4 Inside a pressure-independent valve

Two functions are mounted in series in the same body (R1, R2), with the inlet on the left at p₁, the outlet on the right at p₃ and an intermediate chamber at p₂.

Figure 4: section of a PICV, with the two stages in series. Source: R1, figure 5-2.

The operating range constitutes the limit. Independence from pressure is only ensured between a Δp minimum and a Δp maximum at the valve terminals. Below the threshold, the valve becomes a simple orifice again and the system returns to the coupled regime, without any particular indication.

Two consequences follow. A pressure-independent device consumes its Δp minimum, 25 to 35 kPa depending on the range, drawn from the available manometric height: this hydraulic cost is not negligible. It also contributes nothing in scenarios of pressure loss, specifically those studied in transient conditions, since below the threshold it stops regulating and the flow collapses like with a passive device. Finally, the diaphragm cartridge is a moving part of small section, thus sensitive to fouling.

3. The heating of buildings, the origin of these devices

Each device in the table above was invented for buildings, and it is worth understanding why: these reasons are precisely what a data centre designer must confront in their own case.

The first case is the oldest : a boiler and a set of radiators in parallel on two pipes. The nearby radiators heat, while the distant ones remain cold. The historical fix is the simplest possible application of the branch law: a balancing valve on each radiator return, set once and then forgotten, which corresponds exactly to the DRV of § 2.3. Its modern industrial form, an adjustable orifice with two pressure taps and a preset index, remains the common device of the trade (R4). European practice considers this adjustment as part of the work, not as an optional refinement: EN 14336, the standard on the installation and commissioning of water heating systems, dedicates an annex to good practices for balancing water flows (B1).

Then the terminals began to modulate. Thermostatic valves, then two-way control valves of fan coil units, transformed each branch into a variable resistance controlled by the room. This results in four characteristics that make building hydronics difficult:

• The network is never at a single operating point. The total flow continuously oscillates between almost zero and full load; Δp_dist therefore varies with it and the adjustment made at commissioning is only accurate for a single load.
• The diversity is great. A south facade closes while a north facade opens; the flow taken by part of the building is not correlated to that of the rest.
• The distribution is long and of small diameter. The rising columns are sized just right, so α is high and the coupling is first order, not a correction.
• The excess flow is costly. An excessive flow in a partially loaded emitter returns water that is too hot, which degrades the temperature difference of the system and the efficiency of pumping. In chilled water systems, this is the classic low delta-T syndrome, which Taylor largely attributes the causes and remedies to regulation and balancing defects (B2).

The industry has mobilised the three levers in order : first differential pressure control valves by rising column, which stabilise the Δp applied to a group of terminals (lever 2, applied locally), then pressure-independent balancing valves, which make each terminal indifferent to the pressure applied to it, and then the PICV, which combines balancing, regulation, and shut-off in a single body and eliminates any measurement at commissioning. The idronics 34 from Caleffi presents this evolution family by family (R1). The PICV therefore constitutes the right answer for a network where levers 1 and 2 cannot be pushed far enough.

This context should not be blindly transposed. The demonstration of figure 3, often used to justify pressure-independent valves, assumes a fixed-speed pump supplying modulating terminals. With a pump regulated on Δp and fixed terminal flows, the same figure loses much of its significance.

4. Liquid cooling in data centres

4.1 Why DLC is the favourable case

A DLC network is hydraulically identical to figure 1: a CDU between the facility water system (FWS, Facility Water System) and the technology cooling system (TCS, Technology Cooling System), two headers and a branch per rack including cold plates, hoses and quick connectors. But each of the characteristics from § 3 is reversed.

The last line is decisive. A CDU is by design a regulated source: the specifications require it to provide a « flow control », that is to « regulate and balance coolant flow delivered to racks or cooling manifolds » and to « maintain stable differential pressure under changing IT load conditions » (D6). Translation: regulate and balance the fluid flow delivered to the racks or manifolds, and maintain a stable differential pressure when the IT load varies. The OCP design guide explicitly outlines the resulting strategy:

« By understanding the total pressure drop and estimating the pressure loss within the supply and return pipes, the CDU’s pump speed can be regulated to deliver the required flow rate to each rack at the desired pressure. This target pressure drop serves as the setpoint for the entire system.
Adapting to changes in rack configuration: as racks are added or removed, the pump dynamically adjusts its speed to maintain the set pressure drop, thereby delivering the appropriate flow rate to meet the cooling demands of the connected racks.
Ensuring balanced flow: to prevent unbalanced pressures that may cause some valves to exceed their gauge limits at high flow rates, it is critical to regulate the flow to each rack. This is achieved using flow control valves installed on each branch. »  D2

Translation: by knowing the total pressure drop and that of the supply and return pipes, the speed of the CDU pumps can be regulated to provide each bay with the required flow at the desired pressure. This target pressure drop becomes the setpoint for the entire system. When adding or removing bays, the pump adjusts its speed to maintain this setpoint and provide the corresponding flow. Control valves installed on each branch finally ensure the balancing of flows and prevent some valves from exceeding their pressure limits.

This is, word for word, the diagram of § 1.2: a setpoint Δp at the system level, automatic adaptation to the number of bays, and a branch adjustment device whose role is limited to setting a resistance.

4.2 When a manual valve is sufficient

Three conditions, which are exactly the assumptions of § 1.2.

Once balanced, such a network remains balanced. This is precisely OVHcloud's argument for the manual valve: it is suitable for « large installations with server racks […] where the configuration of the racks to be installed is well-documented over a period », when « the pressure drop is precisely defined based on the cooling circuit within the rack » and « the system can be simulated or tested before the final deployment » (D3). Translation: it is suitable for large installations where the configuration of the racks and the pressure drops are well defined, and which can be simulated or tested before final deployment. OCP cites manual valves as a valid flow adjustment option (D2) and its recommendations for connecting to the racks extend to a configuration « double isolation valve + grooved coupling », consisting of two isolation valves and a grooved coupling. The document specifies that « double ball valves enable use of vent / drain / bypass connections to address commissioning and onboarding challenges as well as flow stabilization » (D2), meaning that two ball valves allow for the necessary purge, drain, or bypass connections for commissioning and flow stabilization. The branch device can therefore be a manual balancing valve of the IMI TA or STAD type (a DRV with pressure taps), or even a marked isolation ball valve. Pressure taps are essential, as they allow the commissioning team to measure and adjust the flow, unlike a pressure-independent cartridge.

4.3 When this is not enough

Four situations invalidate the previous conditions and justify a pressure-independent or intelligent valve:

• Long, branched, or undersized collectors, therefore α greater than 10%. A typical case is a branch in DN 25 (1"): at identical power and bay flow, OCP indicates that its pressure drop can reach five times that of a DN 50 (2") pipe (D2). The distribution then consumes a significant portion of the Δp available and the branches reconnect.
• Large disparity of power between racks on the same distribution, a situation explicitly addressed in the MCV white paper as a branch balancing case (D3).
• Gradual deployment. Between the delivery of the network and the occupation of all locations, the total flow varies greatly. The current solution is not a PICV, but a compensating bypass : OCP recommends the « incorporation of a regulating valve into the bypass connection to impose a false resistance equivalent to that of the rack + manifolds » (D2), that is, a regulating valve on the bypass imposing a resistance equivalent to that of the rack and its collectors. This allows for the validation of the CDU flow and for pre-setting the valves before the actual arrival of the racks.
• Need for telemetry by rack. A manual valve does not transmit any flow measurement. If this flow is to become a monitored point, a flow meter or an intelligent valve becomes necessary.

4.4 What the DLC references say

There is no prescriptive standard for balancing a TCS loop, but a set of now converging guides.

Two cross-cutting observations. First, none of these documents prescribes a type of valve: all refer to a case-by-case analysis, and D2 indique que « the pros and cons of control methodologies (manual, automatic & use of telemetry) is a complex discussion beyond the scope of this paper », soit que les avantages et les inconvénients des méthodes de régulation manuelles, automatiques ou par télémétrie dépassent le cadre du document. Deuxièmement, tous convergent vers la nécessité d’un modèle hydraulique du réseau, que ce soit pour la modélisation transitoire requise par D1, the pre-deployment simulation recommended by D3, or the « transfer of hydraulic models to support future planning » requested during commissioning by D2, that is, the transfer of hydraulic models intended to support future developments.

5. The impact on the modelling of balancing systems

At SIL3X, we specialise in the dynamic modelling of energy systems. The same equations apply for distribution to local cooling units in nuclear and for a data centre DLC loop: a source (CDU or group), a distribution network, parallel branches, and a flow setpoint or Δp to be maintained under variable load.

The right level of model. Several representations are possible. The criterion is not sophistication, it is the question to be addressed. A model that is too simplistic misses the phenomenon (authority, pump switching, thermal front). A model that is too rich, compressible, inertial, meshed beyond necessity, makes the calculation too slow to simulate the entire loop, the redundancies, and the commissioning scenarios. One does not use a hammer to kill a fly: one chooses the smallest set of equations that captures the quantities of interest, in order to maintain numerical performance compatible with the entire network (D1, D2, D3).

Three levels cover the essentials of balancing and regulation studies.

1. Simple thermal, imposed flows. Energy balance on volumes, heat exchangers, and racks, with prescribed flows (nominal, or from a scenario). This is the fastest model: it allows for the representation of the entire loop, IT powers, supply and return temperatures, and the thermal sizing of the CDU. Its limit is structural: without a hydraulic network, there is neither pressure, nor flow sharing, nor transient of Δp. It therefore cannot handle a pump switch, a branch closure, or an authority drift α, nor the effect of pressure losses from the CDU and the collectors on the pressure available at the bays. It remains suitable when the flow rate is an assumption, not an unknown.

2. Hydraulics (quasi-static or slow dynamic). Pumps, characteristics H(Q), valves (K_v), pressure losses from piping, collector and CDU, regulation of Δp or flow rate. This is the model requested by the guides: simulation before deployment (D3), transfer of hydraulic models for developments (D2), criterion α and 1/40 (D4). It becomes necessary as soon as the losses from the CDU, the heat exchanger and the collectors weigh on the pressure at the ends of the branches, thus affecting the actual flow rate delivered. It allows for the sizing of balancing valves, to check the setpoint Δp (D6) and to compare a manual valve to a PICV. Transport is not modelled, which makes the complex transients of electrical loss, chiller shutdowns and recovery by buffer tanks inaccessible.

3. Hydraulics and thermal transport. We add the transport of temperature fronts in the pipes (advection, fluid volumes, delays). This level is justified when the phenomenon of interest is a front (which quickly becomes the case with very large piping networks), and not just a flow rate: pump switching without active redundancy, mixing FWS / TCS, network length, or the temperature jump of TCS up to 27 °C cited by ASHRAE (D1). Without transport, a hydraulic model provides the flow rates immediately, but misses the delay and amplitude seen by the cold plates.

In practice, the three are often combined: imposed flow rates for an initial energy balance of the entire loop, hydraulic for balancing and CDU losses, transport only on the sections where volume and travel time matter. The model remains light enough to cover N+1, the gradual deployment and commissioning, while staying true to the quantities that the guides require to be checked.

5. Takeaways

6. References

Figures 3, 4 and 5 come from issue 34 of idronics published by Caleffi (R1), an open access publication reproduced here with attribution for technical illustration purposes. It is the most comprehensive source on balancing valve families. Figure 2 is in the public domain (B3).

References and guides from the DLC industry

Heating of buildings and hydronics

Manufacturer documentation