BP神经网络优化的无迹卡尔曼滤波核事故源项反演方法研究

On the source nuclide inversion approach to the unscented Kalman filter based on the BP neural network in nuclear accidents

核事故发生后,通常很难根据厂内仪表数据判断事故严重程度,只能通过厂外监测数据估计源项释放量。利用CALPUFF软件模拟了核事故后放射性核素I-131释放的过程。将无迹卡尔曼滤波算法与高斯多烟团大气扩散模型结合,构建无迹卡尔曼滤波核事故源项反演方法模型,实现核事故后厂外源项实时跟踪反演。由于污染物I-131在大气中扩散过程受温度、气压、风速、云量、太阳辐射、地形条件等多种因素共同影响,源项反演为复杂的非线性问题。针对高斯多烟团大气扩散模型将实际条件简化引起的缺陷,使用BP神经网络对反演模型中的量测方程进行优化,从而减小模拟结果与真实情况的误差。无迹卡尔曼滤波核事故源项反演方法模型与卡尔曼滤波核事故源项反演模型相比,能更好地适应非线性条件下释放率变化的情况,反演结果与真实值更接近。

The given paper takes it as its goal tooptimizethe measuring equationof theunscented Kalman filter modelin calculating thenuclear accident source nuclide inversionin theBP neural network. For the said purpose, we have establishedan observation equation with the atmospheric diffusion model of Gaussian multipuff, and built up the unscented Kalman filtersource item inversion model of the nuclear accident in MATLAB programsvia the CALPUFF software. Practically speaking, the source item inversion is by nature belonging to a nonlinear problem, forthe diffusion process of pollutantsI - 131 in the atmosphere tends to beinfluenced by quite a number of factors, such as the temperature, the air pressure, the wind speed, the cloud covering situation, the solar radiation and some other surrounding ground-up conditions. Therefore, the environmental monitoring data should be said including all the above mentioned izffluential factors with the releasing rates and measurement equation matrix elements being part of the output results. According to the comparative analysis of the inversion speeds and the average relative errors between the inversion release rates and the actually existing release rates, research has to be done to optimizethe performance of the Kalman filter nuclide inversion modeland the unscented Kalman filter in the BP neural network system. The results of our research show that the average relative error of the unscented Kalman filter can be reduced by 35.11% ascompared with the Kalman filter model. The inversion data can be madecloser to the real release rate, which is faster and better adapted to the change of the release rates under the nonlinear condition. Under the changing release rates, the inversion outcome of the unscented Kalman filter opti- mized by BP neural network can be predicted to be the best with an average relative error of 24. 34%. The response should bethe fastest or prompt when the accident takes place. Hence, under the condition of the nonlinear release rate changes, the model may enjoy a better adaptability, with the inversion results being closer to the real values when compared with the unscented Kalman filter model.