|
About the author
Chetan Saundankar is the CEO of Coditation, an enterprise engineering and AI services firm working with organisations globally to deliver measurable business impact through intelligent automation, data engineering and platform modernisation.
|
Having worked deep inside power plants with AI, I’ ve seen firsthand the industry’ s drive to innovate. I run Plant360. AI, and much of our work sits inside power generation and industrial energy systems. The model is rarely the hard part. What’ s hard is building something that changes what an actuator, control valve or pump actually does on the floor, not what a dashboard says in a conference room. The real problem is rarely the one in the pitch deck
Most conversations start the same way: Can we forecast solar output better? Cut curtailment? Catch failures earlier? Fair questions, but once we’ re inside the operation, the real issue is usually something else. It’ s almost never the thing in the RFP. It’ s rarely forecast accuracy but the gap between a forecast landing on someone’ s screen and someone acting on it. I’ ve watched teams celebrate a model that beats their old tools by 4 or 5 points, then lose that gain in the two hours it takes to see, question and act on the alert. One engagement drove this home. The client wanted better renewable forecasting. We tightened the model and got competitive numbers. But the real cost was the ninety-odd minutes between“ solar’ s about to drop” and someone adjusting battery charging or starting a flexible generator. Every predicted shortfall now routes straight into a dispatch recommendation, with reasoning and action attached. We didn’ t make the forecast better. We shortened the distance between knowing and doing. That’ s where the money sat. The forecast was never the point. The lead time it buys the equipment that must respond is the point.
How AI Is reshaping power generation and storage
Three patterns keep showing up, project after project. 1. AI as an operating layer, not a monitoring layer What actually moves the needle now feeds directly into operating decisions:
• Predictive maintenance models that go beyond“ anomaly detected” to identify
|
a likely failure mode, remaining life and next step.
• Battery management systems catching temperature drift or voltage gaps between cells early enough to pull a module before a safety event.
• Combustion and boiler tuning models continuously adjusting fuel-to-air ratios and steam pressure instead of leaving it to operator intuition.
• Storage dispatch systems deciding hour by hour whether a battery arbitrages, provides frequency regulation or holds reserve for a potential peak.
That’ s the line: AI that informs versus AI that participates. The IEA estimates savings potential of up to USD 110 billion per year by 2035 across plant operations and maintenance. That value only shows up when AI becomes part of the operating loop. 2. Generic AI does not survive contact with a real plant This has cost some clients real money. A generic anomaly tool that works in a call centre can miss the difference between normal turbine gearbox wear and early blade icing. A forecasting model without local terrain data can underperform when moved somewhere new. Nuclear makes the point sharpest. AI has real value in condition monitoring, outage scheduling and efficiency analysis. It has no business overriding a safety system and shouldn’ t get near a control room without rigorous validation. Don’ t file that under nuclear only. The same discipline applies to something as unglamorous as a cooling-loop control valve.
3. The control layer stays independent, always This is the principle I hold hardest. The AI deployments that work are not autonomous; they advise. They sit above the control layer- PLCs, DCS, EMS, BMS- which keeps running even if the AI stack goes dark. Thermal sensors, fuses, contactors, fire suppression and hardwired protection logic don’ t report to the model. AI’ s job is to catch the pattern early and hand the operator something usable. The
|
π |