中国实验快堆热工参数的自适应BP神经网络预测方法研究

中国实验快堆（CEFR）堆芯的热工参数是否超出限值是评价反应堆安全运行的标准。本文针对燃料包壳最高温度预测问题，通过堆芯子通道分析程序COBRA生成数据样本后，开发基于BP神经网络自适应算法的智能预测程序，对于特定的单盒组件，仅需给出堆芯进口功率和流量，即可实现燃料包壳最高温度的快速准确预测。结果表明，与COBRA相比，在大规模重复性计算的场景下，自开发程序能节约大量计算时间和算力，提高燃料包壳设计和CEFR运行时的操作效率。实验分析得出BP神经网络方法的最大相对误差不超过6%，平均预测相对误差不超过3%，计算效率提升至原程序的300倍，网络模型的预测精度高，且易推广至实验快堆其他参数预测，具有很大的应用前景。

Research on Adaptive BP Neural Network Prediction Method for Thermal Parameters of China Experimental Fast Reactor

Whether the thermal-hydraulic parameters of China Experimental Fast Reactor (CEFR) core exceed the limit is the standard for evaluating the safe operation of the reactor. For the maximum temperature prediction problem of fuel cladding, after generating the data samples by the core sub-channel analysis code COBRA, an intelligent prediction code based on adaptive BP neural network algorithm was developed in the paper. For a specific single-box component, only the core inlet power and mass flow rate were required to achieve fast and accurate prediction of the fuel cladding maximum temperature. Compared with COBRA, in the scenario of large-scale repetitive calculation, self development code can save a lot of calculation time and rescource, and improve the operating efficiency of fuel cladding design and CEFR operation. The experimental analysis shows that the maximum relative error of BP neural network method is less than 6%, the average prediction relative error is less than 3%, and the calculation efficiency is improved to 300 times of the original code. So the prediction accuracy of the network model is high, and self development code is easy to apply to other parameter predictions of the experimental fast reactor.