Abstract: To increase electric power output and extend the refueling cycle, power up-rates and advanced fuel management techniques have been adopted by commercial boiling water reactor (BWR) operators. However, steam leaving the core will easily saturate moisture-separation mechanisms installed in the reactor vessel, causing moisture carryover (MCO) in which the steam is not fully dried before entering the steam line. This issue will cause both greater impact on turbine components and elevated dose level to onsite personnel, which will degrade overall reactor performance.

A data-driven predictive model for MCO in the General Electric Type-4 BWR is built by using machine learning with physics and engineering constraints. Based on data from operational plants and a preliminary feature selected by using physics and engineering analysis, a predictive machine learning MCO model was built by using a neural network, and the features were fine-optimized by using a genetic algorithm. By minimizing the customized cost function, multiple robust and generalizing neural network models were achieved. The best model can predict MCO from an unseen cycle with an MSE of 9.69E-5, which is in agreement with the measurement uncertainty of MCO values. Such accurate prediction of MCO for an upcoming power generating cycle is of great use for commercial BWR operators, as the operational plan can be revised promptly to avoid the undesired high MCO, thus avoiding the elevated dose level to onsite personnel and the greater impact to the turbine components. Such a result could also be used in the maintenance scheduling of cycles currently in progress.