AI in the Chiller Plant Room: Singapore’s Keppel Bay Tower and the Turning Point for Urban Building Intelligence

In the narrative of urban technology, the spotlight usually falls on what can be seen: driverless shuttle buses, smart lampposts on street corners, the giant data screens on the walls of command centers. But a city’s energy consumption structure is often determined by windowless spaces—the chiller plant rooms in basements, the cooling towers on rooftops, the pumps that run continuously on equipment floors.

Singapore’s Keppel Bay Tower became the country’s first commercial building to receive “Super Low Energy” certification, and the key action that drove this result was not replacing the curtain wall or installing photovoltaics, but implementing AI-driven optimization of the chiller plant room. The significance of this matter lies precisely in its “unattractiveness.”

I. The real load of urban decarbonization is hidden in unseen places

In the electricity consumption curve of commercial buildings, the cooling system is usually the largest single load, especially in tropical cities. Year-round high temperature and high humidity mean that cooling is not a seasonal demand but a foundational one.

The problem is that central chiller plants have long operated in a “safe but inefficient” state. Multiple chillers, cooling towers, chilled-water and cooling-water pumps need to coordinate, while load fluctuates continuously with weather, foot traffic, and tenant usage habits. Traditional building management systems (BMS) mostly rely on fixed rules and schedules: set a temperature range, and start/stop units based on experience. To avoid complaints, operating strategies naturally lean toward “overcooling.” This conservatism does not come from technological backwardness, but from the accountability structure—operations and maintenance staff bear the risk of failure, not energy efficiency metrics.

The value of AI entering this step is not to provide new hardware, but to continuously answer a question that traditional rule-based systems struggle to answer: under current load, weather, and equipment conditions, which combination of operating parameters is least costly. This optimization does not require replacing units, nor does it require tenants to make any perceptible compromise, yet it can directly affect a building’s energy baseline and carbon emissions performance.

This is also why cooling optimization occupies a special place in the building decarbonization agenda: it is one of the few areas where “software” can leverage “existing stock.”

II. Certification: Turning “performance” into a governable object

The Singapore Building and Construction Authority’s introduction of the “Super Low Energy” category under the Green Mark framework is the key institutional context for understanding this matter. Traditional green building assessments rely heavily on design-stage simulations and commitments, whereas Super Low Energy certification points to measured performance—whether a building really uses less energy in actual operation.

This shift appears technical, but it is in fact a change in governance logic. Once a city begins certifying “operational performance,” the government is no longer merely setting standards, but becomes a requester and verifier of data. Building owners must continuously collect, report, and verify energy consumption data; certification bodies need comparable measurement standards; and the market needs a signal that can be recognized and priced.In other words, performance certification is essentially a measurement infrastructure. It partially transforms a building from a physical asset into a data object that can be regulated and compared. This is also the most concrete form of the digitalization of public administration in the building sector: not moving documents online, but turning outcomes into real-time verifiable data streams.

III. AI Enters the Operational Layer: The Difficulty Is Not in the Algorithm

Using a machine learning model for chiller plant optimization is not new at the algorithmic level. What is truly difficult is the part before the model: whether sensors are complete, whether metering covers major equipment, whether the data sampling frequency is sufficient, whether BMS protocols from different vendors are interoperable, and whether historical data has large gaps.

Buildings are typical “heterogeneous legacy systems.” A commercial building that has been in operation for years often mixes equipment from different eras, brands, and communication protocols. Data does not exist naturally; it is “wired” out layer by layer. The prerequisite for AI optimization is first establishing complete observability of the chiller plant—this is an engineering problem, not a model problem.

The data center industry is also worth referencing. Google and DeepMind publicly disclosed that after applying machine learning to data center cooling control, cooling energy consumption dropped significantly. But data centers have far higher equipment homogeneity, parameter controllability, and data literacy among O&M teams than ordinary commercial office buildings. Transferring this logic to an urban building often requires not computing power costs, but the costs of instrumentation retrofits, system integration, and O&M process reengineering.

IV. Retrofit, Not New Construction: Existing Buildings Are the Main Battlefield

Most studies on building decarbonization point to the same fact: of the buildings that will still be in use by the middle of this century, the vast majority have already been built today. New projects can serve as demonstrations, but they cannot change the total.

Therefore, the digitalization path for existing buildings usually shows clear phases:

1. Visibility: start with sub-metering and data collection, turning energy consumption from a “month-end bill” into a “minute-level curve”;
2. Diagnosis: identify abnormal operation, equipment degradation, and unreasonable setpoints;
3. Control: gradually introduce predictive control and automated optimization to replace fixed rules;
4. Coordination: enable buildings to respond to grid signals, price signals, and carbon signals.

The significance of the Keppel Bay Tower case is that it validated the feasibility of this path in a commercial office building scenario: rather than relying on tearing down and starting over, it achieves the performance threshold set by the certification system through continuous optimization at the operational layer. This has a direct impact on owners’ decision-making logic—the payback period of retrofits can be calculated, and construction causes limited disruption to tenants.

V. From Individual Buildings to City-Scale Energy Systems

If building intelligence stops at “each building saving electricity on its own,” its value is linear. The truly structural change occurs when buildings begin to be dispatched as grid nodes.Singapore’s large-scale district cooling system deployed around Marina Bay offers a centralized approach: shifting cooling from the rooftops of each building to underground pipe networks and centralized cooling plants. Optimizing an individual building’s cooling plant with AI, by contrast, represents another distributed approach. The two are not opposites; together they form an urban cooling energy system capable of cascaded dispatch.

In tropical cities, cooling load is highly predictable, making it a high-quality flexible resource. Through pre-cooling, thermal storage, and dynamic adjustment of setpoints, buildings can reduce power during grid peak periods and store energy during off-peak periods. When enough buildings have this capability, together they constitute a virtual power plant. At this point, buildings are no longer merely energy consumers, but regulating units of the urban power system.

This is also the most pragmatic landing point for digital twins in the urban energy field: not rendering a beautiful three-dimensional city, but enabling the operating state of every building to be predicted, simulated, and incorporated into dispatch.

6. The Reconstruction of the O&M Profession, and the Overlooked Risk of Dependency

AI entering the cooling plant room changes not only how equipment operates, but also people’s roles. The core competence of traditional O&M engineers is experiential judgment and fault handling; as control logic is gradually taken over by models, the focus shifts to data quality validation, model boundary judgment, anomaly tracing, and cross-system coordination.

This brings two issues that must be faced squarely.

The first is transparency. If optimization logic is encapsulated in a vendor’s black box, owners may obtain lower energy consumption while losing understanding of how their own assets operate. Once a contract ends or the system stops being updated, a building may be unable to return to its original operating state. Auditability, explainability, and exitability should become basic clauses in building AI procurement.

The second is new technological dependency. Building energy efficiency was originally a localized, decentralized capability; after becoming intelligent, it begins to depend on remote platforms, algorithm service providers, and continuous data connectivity. At the city level, this dependency is an efficiency dividend; at the level of an individual asset, it is risk exposure.

7. Buildings Become Digital Infrastructure, and Thus an Attack Surface

When cooling plants, elevators, lighting, and access control are uniformly connected to the network and dispatched by algorithms, a building has in fact become a small industrial control system. While gaining efficiency, it also acquires the security risks inherent to industrial control systems.

Building automation systems being hacked, encrypted by ransomware, or having setpoints tampered with are no longer hypothetical scenarios, but issues of ongoing concern in the global facility management field. For cities, this means buildings are not only an energy issue, but also a cybersecurity issue.

There is also the accompanying question of data ownership: do the energy consumption data, people flow data, and equipment operation data in a building belong to the owner, tenants, equipment vendors, or regulators? In jurisdictions lacking clear rules, such data are often de facto held by platform operators. The issue of digital sovereignty also holds at the building scale.

8. Institutions Are Scarcer Than TechnologyIf we widen our view to the global scale, we can see that different cities have chosen different policy instruments: New York uses carbon emission caps and penalties to constrain large buildings; in its revised Energy Performance of Buildings Directive, the EU is pushing new buildings toward zero emissions and setting minimum energy performance requirements for non-residential buildings; Hong Kong, Tokyo, Dubai, and others have each developed their own certification, disclosure, or benchmarking systems.

These instruments take different forms, yet they share the same premise: regulators must be able to obtain real, comparable, and continuous operational data. In any jurisdiction that lacks measurement capability, rules will degenerate into paper compliance.

By contrast, Singapore’s advantage does not lie in more advanced algorithms, but in the fact that it combines certification standards, data requirements, and industrial capabilities into an executable structure. In this sense, Keppel Bay Tower is a model that was “pushed” out by institutions, not one that was “pulled” out by technology.

Conclusion: A city’s intelligence happens in the basement

Whether a city has truly become smarter can rarely be seen from its skyline. A more reliable observation point is whether the systems already keeping the city running have become better in a measurable sense.

Keppel Bay Tower is just one building. But its direction is worth recording: from showcase to operation, from design to measured performance, from equipment to data, from single-point energy savings to system coordination. The next stage of urban competition is likely to be not about who has more sensors, but about who can turn this data into a continuously operating, auditable, and dispatchable capability.

The starting point of this may be just the chiller in the basement that has finally begun to “think.”