Abstract As the storage terms of spent nuclear fuel (SNF) are extended, monitoring of safety related parameters becomes more important for safe operation of the dry storage casks. The cask integrity assurance will be greatly enhanced if the canister pressure and the peak cladding temperature (PCT) can be reliably inferred from canister surface temperatures (CSTs), and this would also eliminate need for any measurement sensor that penetrates the canister pressure boundary. In the present work, a method to reliably predict the pressure and the PCT using artificial neural network (ANN) models has been developed using CSTs measured at axially spaced positions. To validate the method, a test rig with a single fuel assembly was constructed, and pressure variation tests were carried out. ANN models were developed using the experimental data and the models were augmented by the functional link with CST profile shape information for improved prediction performance. A computational fluid dynamics (CFD) code was then used to simulate these pressure variation tests, and comparison was made to assess the discrepancies between the simulation and the measurement results. The ANN models were trained on the databases built from the CFD calculation results. The measured CST signals were converted to the CSTs of simulation hyperspace on which the ANN models were trained. The conversion was done with the compensation factors obtained at a known reference state. The ANN monitoring models showed good prediction performances in spite of the discrepancies between the measurement and the simulation results.

中文翻译：

---

基于人工神经网络的干式储罐完整性监测方法

摘要 随着乏核燃料（SNF）储存期限的延长，安全相关参数的监测对于干式储存容器的安全运行变得更加重要。如果可以从罐表面温度 (CST) 中可靠地推断出罐压力和峰值包层温度 (PCT)，那么罐的完整性保证将大大增强，这也将消除对任何穿透罐压力边界的测量传感器的需要。在目前的工作中，使用在轴向间隔位置测量的 CST 开发了一种使用人工神经网络 (ANN) 模型可靠地预测压力和 PCT 的方法。为了验证该方法，建造了一个带有单个燃料组件的试验台，并进行了压力变化试验。人工神经网络模型是使用实验数据开发的，模型通过与 CST 轮廓形状信息的功能链接来增强，以提高预测性能。然后使用计算流体动力学 (CFD) 代码来模拟这些压力变化测试，并进行比较以评估模拟和测量结果之间的差异。ANN 模型在根据 CFD 计算结果构建的数据库上进行训练。测量的 CST 信号被转换为模拟超空间的 CST，ANN 模型在其上训练。转换是使用在已知参考状态下获得的补偿因子完成的。尽管测量和模拟结果之间存在差异，但人工神经网络监测模型显示出良好的预测性能。