This paper suggests a machine learning strategy for accident data in nuclear power plants (NPPs) to assess accident conditions and provide informative real-time predictions for the emergency response organization (ERO) of NPP, which adopts time window based sub-models with a restricted sliding range. To verify the feasibility and effectiveness of the strategy, three problems are defined to be solved by machine learning models which predict break size, accident scenario, and core damage time (CDT) of loss of coolant accidents (LOCAs). To prepare learning and test data sets for sub-models, 10,000 LOCA event cases having various break sizes and accident scenarios were calculated and the results were transformed to hypothetical accident data as suitable for machine learning. The learning process of three models was analyzed in terms of the preparation of specific data sets used for learning and test, and the prediction errors of sub-models and their causes. The analysis of three models showed that the predicted break size, scenario, and CDT have sufficient accuracies and informative indications for supporting the ERO’s accident response decision making. For a better understanding of the prediction capabilities of the models and their applicability to real situations, an integrated prediction model (IPM) based on the combination of three models is designed and applied to tens of LOCA simulation cases whose characteristics like break sizes are chosen as suitable for sensitivity analysis purpose. The analysis of IPM prediction for simulation cases shows that overall prediction accuracy is very satisfactory and most errors in the prediction are caused by inherent difficulties in the specific accident condition itself. IPM’s prediction is reliable and understandable in most accident conditions. This concludes that the suggested learning strategy is highly applicable and very effective for the assessment of accident conditions in NPPs by machine learning.

本文提出了一种针对核电厂（NPPs）事故数据的机器学习策略，以评估事故情况并为NPP应急组织（ERO）提供信息丰富的实时预测，该方法采用基于时间窗口的受限子模型滑动范围。为了验证该策略的可行性和有效性，定义了机器学习模型要解决的三个问题，这些模型预测了冷却液事故（LOCA）的断裂尺寸，事故场景和堆芯损坏时间（CDT）。为了准备用于子模型的学习和测试数据集，计算了10,000个具有各种中断大小和事故场景的LOCA事件案例，并将结果转换为适合机器学习的假设事故数据。通过准备用于学习和测试的特定数据集以及子模型的预测误差及其原因，分析了三种模型的学习过程。对三个模型的分析表明，预测的中断大小，场景和CDT具有足够的准确性和信息量，可支持ERO的事故响应决策。为了更好地理解模型的预测能力及其对实际情况的适用性，设计了基于三种模型组合的集成预测模型（IPM），并将其应用于数十种LOCA仿真案例，这些案例的特征如中断大小被选择为适用于敏感性分析目的。对模拟案例的IPM预测分析表明，总体预测准确性非常令人满意，并且预测中的大多数错误是由特定事故情况本身固有的困难引起的。IPM的预测在大多数事故情况下都是可靠且可以理解的。由此得出结论，建议的学习策略对于通过机器学习评估NPP中的事故状况非常适用且非常有效。