YBCO带材直流冲击特性的LM神经网络预测

Predicting Resistance of YBCO Tape under Direct Current Impact Based on LM Neural Network

柔性直流电网中的直流(DC)侧短路故障电流会严重危害电网的运行,而电阻型超导故障限流器(RSFCL)能有效地限制短路故障电流的增长,降低对直流断路器开断容量和开断时间的要求。为了研究用于RSFCL的氧化钇钡铜(YBCO)超导带材在短时直流冲击电流下的电阻特性,根据故障电流特征搭建了高压直流冲击平台,实验测量了在不同电流峰值和不同冲击时间的冲击电流下YBCO带材的电阻变化情况。详细分析了失超电阻越过拐点阻值前后的不同变化趋势并解释了其产生原因。据此,分阶段建立了基于Levenberg-Marquardt(LM)算法的多层前馈神经网络,并利用实验室获得的实验数据对网络进行训练和网络结构的优化。利用训练好的神经网络建立YBCO带材直流冲击特性预测模型。预测结果与实验结果的对比表明,基于LM神经网络的建模方法可以有效地预测直流冲击下YBCO带材失超电阻的变化。所得的预测模型可用于研究RSFCL在柔直电网中的设计与应用。

The direct current(DC)short-circuit fault current endangers the operation of the power gird seriously,while the resistive superconducting fault current limiter(RSFCL)could effectively limit the increase of the short-circuit fault current and reduce the requirements for the breaking capacity and breaking time of the DC breaker.In recent years,with remarkable progress in the preparation technology,yttrium barium copper oxide(YBCO)superconducting tape has become the main material for developing RSFCL at home and abroad.However,it is still difficult to model the resistance characteristics of YBCO tape under short-time DC impact current accurately.Therefore,this paper proposes a modeling method to predict the resistance of YBCO tape under DC current impact based on LM neural network.The study of DC impact characteristics of YBCO tapes and the modeling analysis based on neural network were carried out in this paper.In order to study the resistance characteristics of YBCO tape used in RSFCL under short-time DC impact current,a high-voltage DC impact platform was established according to the fault current characteristics.By adjusting the inductance value,capacitance value,resistance value and capacitor charging voltage,the platform can realize the short-time DC impact process with different current peak values and impact time.The resistance changes of YBCO tape were measured under different impact currents whose current peak at 1000 to 3000 A and impact time at 1.9 to 8 ms.The results would be used in the analysis and establishment of the model.The experimental results showed that when the peak of DC impact current was small,the resistance of YBCO tape increased with the increase of current and decreased to zero with the decrease of current,but when the peak of DC impact current was large,the resistance of YBCO tape did not decrease with the decrease of current,it just kept getting higher.The difference between the two situations was whether or not the resistance curve crosses an inflection point.The inflection point was about 0.043 Ω·m^(-1).According to the theoretical model based on the three possible states for a superconductor,this was because the transition between superconducting state and flux-flow state was easier in the early stage of DC impact.As a result,the change of the resistance was sensitive to the impact current before turning to normal state,and the speed of quench and recovery was very fast.However,the increase was slowed down after turning to normal state.Besides,with the resistance of the superconducting layer increasing,the current gradually transferred from the YBCO layer to the buffer layer,the substrate layer and the stainless steel layer.Therefore,the experimental measurement of the resistance of the tape after crossing the inflection point of 0.043 Ω·m-1 showed the resistance change of the tape except the superconducting layer with the temperature rise.The interaction of electric field,magnetic field and thermal field during the quench process of YBCO superconducting tape was difficult to describe accurately with simple mathematical formula.To establish the accurate prediction model of DC impact characteristics of YBCO tape,the neural network method based on Levenberg-Marquardt(LM)algorithm was used.First,two neural networks were designed separately with analysis result of the experiments to fit the data before and after the inflection point.It was due to the different stages of YBCO tape quenching showed different patterns of change.The LM algorithm was selected as the training algorithm,because its high optimization efficiency.Besides,it had both the local convergence of Gauss Newton method and the global characteristic of gradient descent,which meant the training effect would be better.Next,the parameters of the two networks were trained repeatedly and the frameworks were modified and optimized,including the input layers,output layers,number of hidden layers and other training parameters.After training,the coefficient of determination(R2) of the two neural networks reached 0.9998 and 0.9996,which suggested that this model had fit the experimental data well.Finally,the model of DC impact characteristics of YBCO tape was established by the trained neural networks and was use to predict its resistance and voltage under DC impact.For the case that the YBCO tape’s resistance did not cross the inflection point and for the case that the YBCO tape overshoot resistance did cross the inflection point,the simulation results were verified by the experimental data.The results showed that the predicted voltage error was all less than 5.1 V.The comparison of simulation and experiment suggested that the simulation results were in good agreement with the experimental results under different impact conditions.It could be inferred that,this accurate modeling method based on LM neural network was feasible and accurately reflects the resistance change characteristics of YBCO superconducting tapes under DC impact.This research could be helpful to the study of resistive superconducting fault current limiter.