Improving Moisture Carryover Predictions Using a Neural Network Model - MTA-NF-005
| Administrative Items | |
|---|---|
| Date | 12/15/2020 |
| Functional Area Where Benefits Will Be Realized | Nuclear Fuels
Engineering Radiation Protection |
| Reference Implementation Guidance | 2020 NEI TIP Award Submittal #56; Neural Networks for Moisture Carryover Predictions (ID: 13343896) |
| Industry SME | EPRI – Michelle Mura
Contact: NuclearPlantMod@epri.com |
| Previous Implementation | Please contact EPRI for implementation examples and contacts. |
| Implementation Enablers | N/A |
| SWEEP Score |
|
| Applicability | All BWR reactor types
All geographic regions |
| Keywords | Artificial intelligence (AI); models; nuclear fuels; BWRs; machine learning; predictive analytics; moisture carryover; steam cycle |
| Business Case Analysis Cross-Reference | N/A |
Description
Boiling water reactors (BWRs) utilize steam produced from the reactor to generate power from a turbine. In ideal operation, a moisture separator and steam dryer would capture all of the moisture in the steam before it reaches the turbine. However, some amount of moisture may pass into the turbine in the form of droplets and is known as Moisture Carryover (MCO). This moisture can carry radionuclides (e.g., cobalt-60) from the reactor, spreading contamination and increasing dose rates to balance‑of‑plant systems, which in turn lead to greater costs during a refueling outage. Additionally, high MCO can cause increased erosion at the turbine and internal surfaces of main steam isolation valves (MSIVs) and decreased electrical output from reduced steam quality, affecting both equipment reliability and generation. High MCO also leads to increased fuel costs from unnecessarily high reload fuel quantities (to mitigate high MCO).
This MTA focuses on the use of a neural network model to predict MCO. A neural network is a model that is “trained” using a set of input data to predict output data. The neural network model uses 3‑D core simulator files and core parameters (e.g., fuel bundle power and exit quality of the steam) as inputs to predict MCO. Neural network models can be used within a fuel cycle to predict MCO behavior based on changing operational conditions and for future fuel cycles to predict MCO behavior based on alternate core loading patterns and control‑rod strategies. These models can reduce adverse effects resulting from high MCO by providing information that can help reduce MCO levels within a fuel cycle and by accurately predicting the amount of fuel needed to meet fuel reload design requirements for future cycles.
Benefits
Benefits Estimate
Level 2 – Savings are expected to be up to $5 million based on a two‑year fuel cycle. Extent of savings are based on MCO levels and fuel reload cycle frequency and are therefore plant‑specific. Potential savings are achieved through reductions in dose and reductions in fuel cost. Additional savings are possible from potential outage avoidance.
Benefits Description
- Enhanced core efficiency and reduced maintenance / fuel reload costs. The neural network model can be used to decrease the number of fuel bundles in a fuel reload while still meeting design requirements. Without the model, plants would lack the basis for decreasing fuel reload sizes and would need to purchase greater reload quantities to mitigate high MCO.
- Improved reliability of plant assets due to decreased MCO. High MCO can cause increased erosion at the turbine and internal surfaces of the MSIVs.
- Reduced radiation fields. High MCO levels lead to greater dose rates from radionuclide carryover. The neural network model provides information that can be used during a fuel cycle to reduce MCO levels and its resulting dose rates.
- Potential for avoiding unplanned outages due to high MCO. The neural network model provides a better resource for determining MCO levels relative to operating limits. Without the neural network model, plants may need to shut down to address exceeded MCO limits.
Costs and Schedule
Cost
Level 3 – Initial implementation costs should be less than $1 million. Additional recurring maintenance costs are expected thereafter (less than $100,000 a year). The cost includes software development and set‑up for a plant. Costs can also include a user interface for plant personnel depending on interest.
Schedule
Implementation schedule depends on available data: developing an MCO model requires at least three historical operating cycles. Implementation and planning will take six months to one year if data is available, and several years if data is not available. Three fuel cycles worth of data will take at least six years to gather (with two‑year refueling cycles).
Scope Context
Per site.
Risks
Standard project risks associated with implementing a new modeling method and utilizing a contractor service at nuclear power plants.
Potential risks from over‑training or under‑training the neural network model. Risk can be mitigated by demonstrating accuracy of the model against measured data and testing the robustness of the model against a wide range of conditions.
Plant operational changes made after a model is already developed could potentially bring plant conditions outside of the model’s predictive capability. Risk can be mitigated by iteratively retraining the model.
The neural network model is only as good as its input data. Thus, the input data needs to be consistent: fuel vendors with different core simulator models may produce values inconsistent with each other.