A 420 MW gas-fired power station is losing 380 hours a year to unplanned outages. Its forced outage rate is 4.2%, and its heat rate is 7,850 kJ/kWh-450 kJ/kWh above its 7,400 design figure. In this scenario, each 1% heat-rate improvement is worth an estimated $1.1 million annually.
That makes the opportunity clear: use plant data to spot developing equipment problems earlier, understand what is driving efficiency losses, and make better-informed operating decisions. The aim is not to hand control of a power station to an algorithm. It is to give engineers timely, useful evidence, and a disciplined way to test and sustain improvements.
This is where AI in power generation connects with Lean Six Sigma. AI can help find patterns in complex operational data; DMAIC gives teams a structure for verifying causes, testing changes and controlling results.
Map the AI toolkit to DMAIC
Define: Set the operational target
Begin with a focused problem statement and agreed definitions. For example: reduce forced outages and improve net heat rate without compromising safety, regulatory compliance, emissions limits or dispatch requirements.
Agree how the team will calculate heat rate, forced outage rate, unplanned outage hours and fuel-related savings. Also define the project boundary: which units, operating modes and time periods are in scope?
This matters because a model can find patterns, but the team must decide whether those patterns relate to a meaningful operational problem.
Measure: Build a reliable view of plant performance
The Measure phase establishes whether the data can support sound analysis. Relevant sources may include the SCADA historian, vibration sensors on turbine bearings, thermography, maintenance records, fuel measurements and operating logs.
Before modelling, engineers should check:
- Tag definitions, units, sampling intervals and data-quality flags
- Sensor calibration and measurement-system reliability
- Differences between start-up, shutdown, transient and steady-state operation
- Whether heat-rate calculations are consistent and adjusted for relevant operating conditions
The result is a trusted baseline, not simply a large dataset. If measurements are inconsistent, AI may identify patterns that reflect instrumentation issues rather than equipment behaviour.
Analyse: Identify and validate likely causes
In the scenario, machine-learning analysis of eight years of operating data finds that 44% of forced outages trace to HRSG boiler tube leaks, with acoustic signatures appearing weeks before conventional indicators. The analysis also finds that heat-rate degradation is associated with condenser-fouling patterns.
Those findings are leads for investigation, not proof of cause. The team should compare model alerts with inspection results, maintenance records, operating conditions and engineering knowledge. Statistical and visual tools (such as trend analysis, stratification and cause-and-effect analysis) help test whether the suspected relationships hold up.
Acoustic monitoring has been used to identify HRSG tube leaks, including cases where signals provided advance warning. However, detection performance depends on sensor placement, site conditions and validation at the individual plant. See the Combined Cycle Journal’s report on acoustic monitoring and HRSG tube leaks and the research study on data-driven boiler tube leak detection and classification.
Improve: Test changes before applying them
A digital twin of the steam cycle can help engineers compare operating regimes and test combustion setpoints in a modelled environment. The team can evaluate potential heat-rate improvements while respecting equipment limits, emissions requirements and operating constraints.
The AI recommendation should then be assessed by the people accountable for safe, reliable operations. A controlled trial, with a defined comparison period and agreed measures, helps distinguish a genuine improvement from changes caused by dispatch, ambient conditions, fuel or maintenance.
Control: Make useful signals part of daily work
Once a change is validated, the Control phase defines how the plant will sustain it. In this scenario, the team monitors daily heat-rate anomaly scores, uses bearing remaining-useful-life estimates to prioritise inspection, and applies pre-agreed thresholds that trigger engineering review.
A useful alert should tell the team what asset is involved, what has changed, how confident the model is and what evidence needs checking. It should support a decision, not bypass established operating authority.
Worked example: From early warnings to measurable gains
After the plant establishes its baseline, the team runs a six-month ML project alongside engineering review and validation. The model identifies a developing bearing fault six weeks before failure, allowing the maintenance team to plan an intervention rather than respond to an unplanned event.
The scenario’s results are:
| Measure | Baseline | After improvement |
|---|---|---|
| Forced outage rate | 4.2% | 2.8% |
| Unplanned outage hours | 380 per year | 250 per year |
| Heat rate | 7,850 kJ/kWh | 7,610 kJ/kWh |
The heat-rate reduction is 240 kJ/kWh, or approximately 3.1% against the baseline. At the scenario’s stated value of $1.1 million per 1% improvement, that equates to about $3.4 million in annual fuel savings.
These figures are an illustrative plant case, not a guaranteed outcome. The estimate assumes comparable generation and fuel-cost conditions. The team should calculate savings against a matched, adjusted baseline and report heat-rate savings separately from the operational value of fewer outage hours to avoid double counting.
Account for real project costs and risks
An honest business case includes more than software or model development. Likely cost areas include:
- Integrating and preparing historian data
- Retrofitting or improving sensors where coverage is insufficient
- Engineering, operations and data-science time over the six-month project
- Model testing across relevant operating modes
- Independent or OEM validation before relying on model outputs
The plant should also plan for the ways models can lose reliability:
- Rare events: A model cannot learn patterns it has never seen. Rare transients and unusual failure modes need engineering safeguards and explicit escalation procedures.
- Changing dispatch: Different loads and operating envelopes can shift model inputs. Models need to recognise operating context and be monitored for drift.
- Unvalidated control changes: Regulatory and safety obligations require engineering review. Keep AI recommendations advisory until they have been rigorously validated and authorised.
- Cybersecurity: Protect operational technology through utility-grade security practices. Define access, data flows and boundaries between analytics environments and plant control systems.
Connect model insights to continuous improvement with Ci Flow
Finding an anomaly is only the beginning. Teams also need to assign investigation, document evidence, record decisions and monitor whether a verified change holds.
Ci Flow can serve as the workflow layer connecting AI insights to daily improvement work. Its digital process workspace supports structured improvement activity; DMAIC project tracking helps teams organise progress and evidence; and AI-assisted root-cause suggestions can help frame investigation. KPI boards can keep agreed performance measures visible for team review.
For a power-generation project, that might mean turning a validated bearing alert into an owned investigation, recording inspection findings, tracking the improvement actions and reviewing heat rate and outage measures through the control plan. The plant’s engineers remain responsible for confirming causes, approving changes and meeting safety, compliance and cybersecurity requirements.
Make AI an engineering-supported improvement capability
AI can strengthen predictive maintenance in a power plant when it is built on dependable measurements, checked against engineering evidence and integrated into a disciplined improvement method. In this scenario, earlier bearing warnings, focused HRSG leak analysis and heat-rate optimisation work together to improve reliability and efficiency.
Explore Ci Flow to connect improvement insights with structured project work, and build your capability with Lean 6 Sigma Hub’s CSSC-accredited Lean Six Sigma training.
Kaizen. Kai-Care. Kai-Done. Lean Six Sigma
