Eliminate Thermal Limit Bias in BWRs with ThermalLimits.ai
ThermalLimits.ai – A prediction platform for online Thermal Limits in BWRs
ThermalLimits.ai is a state-of-the-art tool that yields real world, high value results via Machine Learning. It enables powerful predictive capability into crucial operating limits – ensuring compliance with technical specifications, enabling reduced reload fuel costs, and eliminating operational challenges.
Blue Wave AI Labs is pioneering online thermal limit prediction capability with the creation of ThermalLimits.ai to address these deficiencies. Our proprietary physics-informed approach uses machine learning (ML) to leverage historical fuel cycle data, outputs from core simulators, and past online thermal performance to construct a relia-ble offline surrogate to replace the online feed-back provided by the in-core instrumentation. The underlying methodology used to develop these models is the notion of an error-correction deep neural network that leverages the offline nodal or bundle thermal limit array in conjunction with additional offline datasets to train a network that predicts the corresponding online thermal limit nodal or bundle array, thereby enabling pre-dictive capability for online thermal limits from of-fline nodal power distributions.
Accurate predictions of core-wide and local be-havior are crucial to assuring that targeted mar-gins to the operating limits are maintained. Devia-tion between measured performance and design predictions can lead to operational challenges, such as unplanned derated conditions, premature coastdown, or increased fuel costs by loading more fuel than required for targeted energy pro-duction. Historically, inability to accurately predict online thermal limits from offline methods has challenged core design and cycle management.
While actual operations may at times depart from cycle design basis projections, there exists an in-herent bias between offline and online methods that stems from the nature of the two systems. Both methodologies rely on a three-dimensional neutronics simulator model to calculate the reac-tor’s power, moderator, void, and flow distribu-tions—from which margin to thermal limits can be determined. However, these calculations are approximations, and the offline quantities deter-mined from them are inexact estimates that lead to uncertainty in thermal limits. Online methods, on the other hand, employ an adaptive process through feedback directly from in-core nuclear in-strumentation while the reactor is online.
Up until recently, there has been no reliable method to bridge the gap between online and of-fline methods leading to inaccurate and incon-sistent predictions of online thermal limits.
The predictive capability of ThermalLimits.ai is illustrated above for a typical test cycle for a large BWR. Individual models for each of the MFLPD, MAPRAT, and MFLCPR distributions demonstrate an average reduction in the observed bias by 73% (3.64x) for MFLPD, 46% (1.82x) for MFLCPR, and 67% (3x) for MAPRAT. More-over, across all fuel cycles independently tested, the maximum bias between online values and model predictions never exceeds 3.9% (for MAPRAT and MFLPD) and 1.5% for MFLCPR.
ThermalLimits.ai is a robust state-of-the-art SaaS application for the nuclear power industry that provides unparalleled accuracy for online thermal limit forecasting in both reload core design and cy-cle management engineering applications. Addi-tional capabilities of these models include: An average bias of less than 1% for the max MFLPD and MAPRAT, compared to more than a 4% mean bias for conventional offline methods. – An average bias of less than 0.39% for MFLCPR, representing a 46% reduction in the bias from conventional offline methods. – A 2x reduction in mean bias for MFLCPR com-pared to conventional methods. – More accurate predictions of the full nodal dis-tributions for MAPRAT and MFLPD, reducing the average nodal bias by more than a factor of two. – Correct identification of the online most limiting node/bundle location more than 85% of the time, compared to 60% from off-line methods. – No degradation in model performance when used for mixed cores or during fuel transitions. With these predictive capabilities, BWR operators can design the most economical and efficient re-load cores by eliminating excess design margin, re-ducing rework, and avoiding operational chal-lenges that often result in power de-rates or in-creased coastdown lengths.