实时工况仿真场景下蒸汽发生器液面智能预测模型的实现与验证

The Design and Implementation of an LSTM-Based Steam Generator Level Prediction Model

本文使用基于深度学习方法的长短期记忆(Long Short-Term Memory,LSTM)模型来预测核电站蒸汽发生器的液面工况,以有效地解决蒸汽发生器液面控制过程中的参数预测问题。该模型在SIMULINK环境下,实现并且构建了一个使用传统PID控制器的实时工况验证平台,对其进行功能和性能验证。PID控制回路的反馈信号为LSTM模型提供实时在线的输入更新,从而获得LSTM模型对于液面变化的实时预测结果。通过与传统RNN模型的对比实验证明,LSTM模型所具有的长时间序列数据处理优势对于蒸汽发生器核心参数的预测非常有效而且精准。LSTM模型对液面的预测误差可低至-1.1887×10-4,预测过程的损失值可低至1.4130×10-8。而且实验结果显示,LSTM模型在预测过程中具有较快的收敛能力,可以在更短的时间内输出准确的预测结果,这对于及时而准确地判断蒸汽发生器的工作状况,维持核电站正常运行具有重要意义。利用LSTM模型的快速准确预测能力,还可以在未来对传统的PID控制方法实行智能化改造,提升蒸汽发生器液面控制的快速响应能力。

The Long-Short Term Memory(LSTM)model is applied to the Steam Generator(SG)water level prediction in this work.The model is designed and implemented within the SIMULINK environment,where a real-time validation platform is also constructed using traditional PID controller.The SG water level feedback signal is fed to both the LSTM model and the PID controller,allowing the prediction to be generated online and compared to the actual controlled value.The results have demonstrated the functionality and advantages of the LSTM model,such as the high accuracy with prediction error of-1.1887×10-4,low loss value with MSE of 1.4130×10-8,and quick convergence during the simulation.Discoveries found in this work could enable future exploration of developing deep-learning-based control strategy for nuclear power plants.