Integrating spatial prediction and Bayesian optimization for dynamic event tree optimization branch search

Abstract

The dynamic event tree (DET) method has been extensively utilized in safety analyses of nuclear power plant accidents. However, due to the complex nature of accident response processes, ensuring a comprehensive simulation of accident outcomes and calculating precise failure probability values requires an enormous number of discrete branches generated by DET, which incurs a prohibitive computational cost. This paper proposes a DET branch search method that integrates spatial prediction and Bayesian optimization. This method uses a limited number of DET branch calculation results as prior data and employs Bayesian optimization to automatically search for the optimal branch step size. By integrating Dynamic Time Warping (DTW) classification with data stability testing, the status of unknown spaces is preliminarily determined, gradually reducing the scope of spaces with potential failure risks. Through the selection of critical yet minimal DET branches for a comprehensive simulation of accident outcomes, computational efficiency is enhanced, and accurate failure probability results are achieved. Using the operation time of the turbine-driven auxiliary feedwater system and the main pump shaft seal leakage time in a station blackout (SBO) accident of a CPR1000 nuclear power plant as an example, a DET model is constructed. Ultimately, the computational time and failure probability of the DET branching optimization search method are compared with those of the equally spaced variable values method. The relative error of the failure probability is within 1.96%, while computational efficiency is improved by a factor of 4.82. This approach not only ensures computational accuracy but also enhances the efficiency of DET calculations, offering substantial methodological support for the practical application of DET.