The Problem with Traditional Cooling O&M
Cooling system O&M has long depended on individual operator experience — the industry's veteran troubleshooter who reads spray pattern off a tower, hears a pump bearing in its running note, and feels line vibration by hand. That knowledge is real, but it does not scale, it retires with the person, and it is unavailable at 3 a.m.
The structural problems are delayed response, high labour cost and weak preventive capability. Faults are typically found after efficiency has already degraded or equipment has already been damaged.
Enabling Conditions
Three trends have made a data-driven alternative practical. IoT sensor cost has fallen enough to instrument a loop comprehensively. Edge computing now provides usable compute at the cabinet, so preprocessing and inference can run locally without round-tripping to the cloud. And industrial AI methods have matured to the point where anomaly detection and remaining-useful-life estimation work on real plant data.
Architecture
A practical architecture runs in layers. Instrumentation captures temperature, pressure, flow, conductivity, pH, turbidity, corrosion rate and vibration. Edge gateways aggregate fieldbus traffic, normalise it and run first-pass anomaly detection locally. A cloud or on-premise platform stores the time series, trains models and serves dashboards; a digital twin provides a physics-informed reference against which measured behaviour can be compared.
Monitoring Parameters and Fault Signatures
The informative measurements are the approach temperature across heat exchangers, pump and fan power draw against flow delivered, make-up and blowdown ratio, and vibration spectra on rotating equipment. Each has characteristic signatures: fouling shows as a rising approach at constant load; impeller wear shows as falling flow at constant power; bearing degradation shows as spectral energy migrating to characteristic frequencies.
Prediction Models
Model choice should follow data availability. Where labelled failure history exists, supervised models can classify fault modes directly. Where it does not — the common case — unsupervised anomaly detection on reconstruction error or forecast deviation is the practical starting point, since it needs only normal-operation data.
Physics-informed hybrids tend to generalise best: use the heat-balance and fouling-growth equations as a structural prior, and let the data fit the residual. This keeps predictions physically plausible outside the range of observed data.
Closed-Loop Control and Economic Value
Prediction only pays when it changes an action. Coupling predictions to setpoints — fan speed, pump frequency, blowdown rate, dosing rate — converts forecasts into avoided energy, avoided downtime and avoided chemical use. Value should be measured on those outcomes, not on model accuracy in isolation.
Implementation Path and Caveats
A staged rollout works better than a big-bang deployment: instrument first, collect a clean baseline across a full seasonal cycle, deploy anomaly detection, then add prediction and finally closed-loop control.
Two caveats matter. Data quality dominates model choice — a badly installed sensor will defeat a good algorithm. And predictions should be presented as ranked, explainable recommendations to operators, not as autonomous commands, at least until trust and validation history exist.