Central Chilled Water Plant (CCWP)

• CCWP is a centralized cooling system that uses chillers, pumps, cooling towers, and thermal energy storage to produce and distribute chilled water effectively.
• The system employs supervisory optimization methods that balance pump speeds, chiller loads, and fan power to minimize energy consumption while meeting cooling demands.
• Advanced digital twins and integrated TES enable precise performance tuning and scalability, enhancing energy savings in district and data-center cooling applications.

A Central Chilled Water Plant (CCWP) is a centralized chilled-water production and distribution system in which chillers generate chilled water, chilled-water pumps circulate it to building air-handling units (AHUs) or other cooling loads, condenser-water pumps circulate water on the heat-rejection side, and cooling towers reject heat to ambient. In the literature considered here, the term also encompasses plant-level thermal energy storage (TES), supervisory optimization layers that sit above existing building automation or plant control systems, and, in some high-density computing facilities, decoupled heat-exchanger subloops that are analogous to central plant distribution branches (Vu et al., 2018, Khalil et al., 2016, Guo et al., 21 Aug 2025).

1. Core architecture and plant boundary

A standard CCWP comprises chillers, chilled-water pumps, condenser-water pumps, and cooling towers. On the building-facing side, chilled water leaves the chillers cold, serves buildings or AHUs, absorbs cooling load, and returns warmer. On the heat-rejection side, condenser water leaves the chillers hot, rejects heat at the cooling towers, and returns cooler to the chillers. The chillers couple these two loops. In campus HVAC language, the central plant produces chilled water and steam that are consumed by building AHUs, while the chilled-water side and steam side are independent systems that do not share machinery (Vu et al., 2018, Khalil et al., 2016).

A control-oriented CCWP model can include an aggregate cooling coil, a number of heterogeneous chillers and cooling towers, and a chilled water-based thermal energy storage system. One such model explicitly assumes a primary-secondary pumping strategy, with multiple water-cooled vapor-compression chillers, multiple evaporative cooling towers, chilled-water pumps, cooling-water pumps, and plant-level interconnections forming a chilled-water loop and a cooling-water loop (Guo et al., 21 Aug 2025).

Subsystem	Function
Chillers	Generate chilled water and couple the load side to the heat-rejection side
CHWP	Circulate chilled water to buildings or AHUs
CWP	Circulate condenser water through chillers and towers
Cooling towers	Reject heat to ambient
TES	Shift cooling production across time

The same central-plant logic appears in district-scale and data-center-adjacent systems, but with different boundaries. A campus district cooling system can behave like a centrally managed chilled water plant even when it uses five chiller stations and a common network, while high-density computing plants may replace conventional building coils with coolant distribution units (CDUs) or plate heat exchangers and still preserve the core hydronic logic of central production, distribution, and heat rejection (Huylo et al., 2024, Jadhav et al., 15 May 2026).

2. Thermal and hydraulic operating principles

The operational objective of a CCWP is typically to minimize total plant power while maintaining acceptable plant operation and meeting the required cooling load. One explicit supervisory formulation minimizes the sum of predicted powers of all major plant components,

$$\underset{x}{\text{minimize} \ } J(x) = \sum_i CH_i(x) + \sum_i CHWP_i(x) + \sum_i CWP_i(x) + \sum_i CT_i(x)$$

subject to bounds on control variables and predicted internal operating conditions, with

$x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$

In CCWP terms, this is a supervisory optimal setpoint problem over condenser-water pump speed, chilled-water pump speed, and cooling tower fan speed, under a fixed equipment staging schedule (Vu et al., 2018).

The basic thermal relation used repeatedly in plant surrogates and digital twins is the branch or loop energy balance,

$Q = \dot m c_p \Delta T,$

or, in branch-temperature form,

$$T_{\mathrm{ret},i}(t) = T_{\mathrm{sup}}(t) + \frac{Q_i(t)}{f_i \cdot \dot m(t) \cdot c_p}.$$

This form appears directly in hydronic optimization for parallel subloops and is the plant-level analogue of chilled-water load transport through a CCWP (Jadhav et al., 1 Mar 2026, Jadhav et al., 15 May 2026).

A central operational tradeoff is that more pumping can increase flow and help meet load with warmer water, while more chiller or tower effort can reduce temperatures but consume more compressor and fan power. The optimizer therefore searches for the best balance across pumps, chillers, and towers rather than minimizing any one subsystem in isolation (Vu et al., 2018, Yu et al., 14 May 2026).

For variable-speed pumps and fans, the dominant physical heuristic is the affinity law,

$\frac{P_1}{P_2} = \left(\frac{S_1}{S_2}\right)^3,$

which motivates cubic or low-order polynomial power models for chilled-water pumps, condenser-water pumps, and cooling tower fans. This same cubic dependence also underlies pump-flow optimization in digital-twin studies, although sensitivity analyses have also examined exponents $n \in \{2.0, 2.5, 3.0\}$ to reflect static head, valve effects, or motor efficiency variations (Vu et al., 2018, Jadhav et al., 1 Mar 2026, Yu et al., 14 May 2026).

A common misconception is that CCWP optimization is only a chiller-sequencing problem. Several studies instead separate macro-control from micro-control. One plant-level data-driven study explicitly does not focus on macro-control or staging decisions and treats binary equipment on/off status as a known operating condition; its main real-time manipulated variables are pump and fan speeds, with updates every 2–3 minutes (Vu et al., 2018). By contrast, predictive-control work on multi-chiller plants shows that discrete chiller staging and continuous setpoints can also be optimized jointly, but that doing so introduces a mixed-integer optimal control problem (Boldocký et al., 30 Mar 2026).

3. Supervisory optimization and control methodologies

A major research direction models the CCWP as a decomposed supervisory control problem rather than a monolithic black box. In one industrial deployment, the plant is decomposed into a condenser-water-pump power model, chilled-water-pump power model, cooling-tower power model, chilled-water flow model, condenser-water flow model, condenser-water temperature model, and chiller power model. Single-input single-output modules such as pump and tower power are modeled by polynomial regression, while multiple-input single-output modules such as flow, temperature, and chiller power use a small multilayer perceptron with one hidden layer of 3 hidden nodes and logistic activation (Vu et al., 2018).

That decomposition is tied to a practical online loop: read current plant state and weather through BACnet, predict flow and temperature states, predict subsystem powers, sum them into total plant power, solve the constrained optimization over $x=\{cwp\_speed, chwp\_speed, ct\_speed\}$, and push recommended setpoints every 2–3 minutes. The deployment ran on an industrial PC connected over BACnet to the plant automation system, and the reported real-world savings were more than 7% in the abstract and between 5% to 10% on average in the deployment section (Vu et al., 2018).

An important result from that work is that simple, domain-aligned models outperformed LSTM on most subsystem tasks. For total plant power prediction, the decomposed method achieved 1.86% MAPE, compared with 2.25% for black-box LSTM; for subsystem models, polynomial regression and small MLPs also outperformed LSTM across chilled-water pump, condenser-water pump, cooling tower, chilled-water flow, condenser-water flow, condenser-water temperature, and average chiller power tasks (Vu et al., 2018). This directly contradicts the common assumption that deep recurrent models are automatically superior for CCWP optimization.

Another supervisory approach uses relay-based extremum-seeking control (ESC) to optimize cooling tower fan speed directly against total plant power,

$$J = P_{CWPump} + P_{CHWPump} + P_{Chiller} + P_{CoolTower} + P_{AHU}.$$

In simulation using a Modelica Buildings Library plant exported as an FMU and co-simulated with Python, the method produced average savings of 15.81% in Pasco, WA and 14.47% in Houston, TX versus fixed 100% fan speed during a July cooling-season week. Against an idealized PID controller holding cooling tower leaving temperature at 25°C, the advantage was climate-sensitive: 0.30% in Pasco and 9.85% in Houston (Yu et al., 14 May 2026).

The same ESC study introduced a virtual power meter (VPM) so that a supervisory optimizer can operate without physical power meters. Validation against a real chiller power meter yielded $R^2 = 0.9611$, RMSE = 5.69 kW, and NRMSE = 5.11%, with the central claim that trend fidelity matters more than perfect absolute accuracy for gradient-following optimization (Yu et al., 14 May 2026).

A third line of work addresses the discrete-continuous nature of plant control. Mixed-Integer Nonlinear Differentiable Predictive Control (MI-DPC) trains neural policies offline to approximate mixed-integer model predictive control for parallel multi-chiller plants. In 7-day simulations with two or three chillers, MI-DPC reduced total energy by about 8–11% relative to a rule-based controller while maintaining inference time near $1.9\times 10^{-4}$ s, whereas MI-MPC runtime was capped at 180 s, equal to the sampling period (Boldocký et al., 30 Mar 2026). This suggests that explicit predictive policies can make integrated staging and setpoint control computationally tractable, although the underlying study omits condenser-water pumps, cooling towers, and TES (Boldocký et al., 30 Mar 2026).

Across these methodologies, several practical themes recur: active data enrichment through safe random setpoint perturbations is necessary when historical BAS data cover only a narrow operating envelope; RANSAC or equivalent preprocessing is needed because plants are “not stable systems” and outliers are common; and low-level operating constraints such as flow, temperature, and manufacturer bounds remain indispensable even in data-driven or learning-based frameworks (Vu et al., 2018).

4. Thermal energy storage, peak shaving, and grid interaction

TES extends the CCWP from a real-time cooling plant to an intertemporal scheduling problem. In the University of Texas at Austin district cooling system, the central plant is integrated with a 65 MW CHP plant, a 45,000 ton district cooling system, and two chilled water storage tanks modeled as one equivalent TES with maximum stored energy 175.6 MWh thermal, approximately 50,000 ton-hours, and maximum charging/discharging rate 31.7 MW thermal, approximately 9,000 tons (Huylo et al., 2024).

The cooling-side model uses a lumped whole-plant COP regression,

$x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$0

with

$x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$1

and the TES state update

$x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$2

Over two 72-hour summer periods, optimized TES dispatch provided an additional 3–5 MW of peak reduction, roughly 5–8% of peak campus electric load, and nearly eliminated or completely eliminated use of the supplementary peaking steam turbine on several days (Huylo et al., 2024).

A related data-center study examined chilled water storage in energy and regulation markets using a three-stage MPC framework: baseline power scheduling, regulation reserve scheduling, and real-time power tracking. Over two days, the framework reduced operational costs up to 8.8% ($x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$338.7) energy cost reduction, 6.5$x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$4338.3) from regulation revenues. The TES-enabled system could bid 242–378 kW of regulation capacity, corresponding to 9% to 14% of system nominal power (Fu et al., 2020). Although this study relies partly on server-side flexibility, its plant-side structure—chiller modulation, TES charging/discharging, and receding-horizon scheduling—is directly transferable to TES-assisted CCWPs.

TES also changes the chilled-water reset problem. Conventional TES operation often maintains conservatively low chilled water temperatures throughout the cooling season, but an integrated framework combining relative humidity prediction, dynamic cooling load estimation, cooling coil performance prediction, and TES discharge temperature prediction found that the daily initial TES charging temperature could be increased by an average of 2.55°C compared to fixed-temperature operation, while still maintaining indoor setpoint temperatures. The same study reports that chiller or heat-pump COP improves by about 2–4% per 1°C increase in chilled water supply temperature, implying approximately 5–10% improvement for a 2.55°C increase, and found an average potential chilled-water supply temperature increase of 4.52°C when considering the coil thermal limit alone (Oh et al., 16 Jan 2026).

These TES studies show that storage is not only a load-shifting device. It is also a supervisory degree of freedom for demand limiting, reserve provision, and supply-temperature optimization, provided that coil capacity, humidity control, and end-of-day discharge temperature remain feasible (Huylo et al., 2024, Fu et al., 2020, Oh et al., 16 Jan 2026).

5. High-temperature heat recovery and supplemental chilled-water production

A nontraditional but CCWP-relevant architecture reverses the usual logic of compressor-driven chilled-water production. In the iDataCool system, an IBM iDataPlex dx360 M3 cluster with 3 racks and 72 compute nodes per rack replaced its air-cooling solution with a custom direct water-cooling assembly. The system operates with rack outlet temperatures up to 70°C / 158°F and transfers waste heat through a heat exchanger to the driving circuit of an InvenSor LTC 09 adsorption chiller, which then produces chilled water for a small GPU cluster with 12 kW peak power (Meyer et al., 2013).

The installation consists of five hydraulic circuits: a central cooling circuit using existing chilled water at about 8°C / 46°F, a primary cooling circuit, the rack cooling circuit, the adsorption chiller driving circuit, and a recooling circuit connected to a fan-driven dry recooler. A 3-way valve governed by a PID controller continuously regulates how much rack heat is sent either to the chiller driving circuit or to the primary cooling circuit as auxiliary heat rejection so as to maintain a constant rack inlet temperature (Meyer et al., 2013).

This system is not a conventional CCWP and is explicitly described as a heat-recovery augmentation layer rather than a replacement for the central plant. The adsorption chiller works efficiently already at driving temperatures of around 65°C / 149°F and is in standby below 55°C / 131°F. From 57°C / 135°F to 70°C / 158°F, its COP increases by 90%. The fraction of electric input that could be reused as chilled water is approximately on the order of 25% for 60–70°C / 140–158°F, and with better thermal insulation this could rise by almost a factor of two at 70°C, i.e. close to 50% (Meyer et al., 2013).

The principal limitation is thermal loss to ambient. As rack temperatures rise, more heat escapes to room air, so the fraction of electrical power captured in water decreases with temperature. The paper therefore treats insulation as a first-order design variable and retains the conventional central chilled-water system for backup and trim service (Meyer et al., 2013). A plausible implication is that CCWP evolution in data centers may increasingly include hybrid architectures in which some loads become thermal sources for secondary chilled-water production rather than pure cooling sinks.

6. Digital twins, modeling validity, and plant design co-optimization

The recent literature increasingly treats the CCWP as a digital-twin and co-design problem rather than only a controls problem. One explicit CCWP model for learning-based controllers describes a framework consisting of an aggregate cooling coil, a number of heterogeneous chillers and cooling towers, and a chilled water-based thermal energy storage system. Its distinguishing feature is a constrained optimization-based framework that ensures the cooling coil, chiller, and cooling tower models respect heat-exchanger capacities irrespective of the inputs provided. The overall plant is implemented in Matlab, uses CasADi and IPOPT for embedded nonlinear programs, and is publicly available (Guo et al., 21 Aug 2025).

That wider validity is essential because standard open-literature component models can become grossly erroneous outside nominal operating ranges. The paper’s motivating example shows a conventional Electric:EIR chiller predicting condenser water return temperature near $x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$5 when condenser water flow is reduced to one tenth of nominal, whereas the saturation-aware reformulation constrains the condenser side and instead yields unmet cooling and warmer chilled-water supply, which is physically meaningful for controller training (Guo et al., 21 Aug 2025).

Digital-twin optimization has also been used to quantify the gap between theoretical and deployable CCWP savings. For the Frontier supercomputer cooling system, a validated surrogate based on a Modelica digital twin and one full calendar year of 10-minute data achieved branch-level CV-RMSE below 2.7% and NMBE within 2.5%. Using that surrogate, analytical flow-only optimization achieved 20.4% total energy saving, unconstrained joint optimization of flow rate and supply temperature achieved 30.1%, and ramp-constrained optimization enforcing actuator rate limits achieved 27.8%, preserving 92.4% of theoretical savings (Jadhav et al., 1 Mar 2026).

Related work on Frontier’s parallel-subloop architecture shows that plant topology and software optimization can substitute for each other to a surprising degree. One annual study evaluated all 611 feasible partitions of 25 CDUs into two through six subloops and found that the globally optimal design was a two-subloop plant achieving 35.48% annual cooling energy savings, only 0.18% above the current three-subloop Frontier design at 35.30%. More importantly, dynamic flow fraction optimization reduced design sensitivity by 93%, from a 96 MWh/yr spread under fixed proportional flow to 7 MWh/yr under optimized flow fractions (Jadhav et al., 15 May 2026).

A parallel life-cycle study using operational energy, embodied carbon, and expected unplanned downtime evaluated the same 611 partitions and found a cost-and-carbon optimum at two subloops holding 14 and 11 units, achieving 3,320.7 tonnes of carbon dioxide equivalent and $x = \{cwp\_speed, chwp\_speed, ct\_speed\}.$6100,000 compared to the built four subloop configuration. However, when reliability is treated as a hard constraint set by operations policy, the built four-subloop configuration is consistent with the constrained optimum (Jadhav et al., 13 Jun 2026). This directly challenges the idea that more hydraulic segmentation is always better: the operational-energy benefit can be nearly insensitive to added subloops, while piping embodied carbon and cost increase with loop count.

Taken together, these studies define an emerging CCWP design doctrine. First, plant models used for optimization or reinforcement learning must remain physically meaningful outside normal operating conditions (Guo et al., 21 Aug 2025). Second, coupled optimization of flow, supply temperature, and branch allocation usually outperforms single-variable reset (Jadhav et al., 1 Mar 2026, Jadhav et al., 15 May 2026). Third, software-level supervisory optimization can capture much of the benefit that might otherwise be sought through physical reconfiguration, although redundancy and maintainability constraints can still justify more conservative topologies (Jadhav et al., 13 Jun 2026).