摘要： 构建了磁絮凝处理矿井水实验装置，并使用RS485接口经串口协议与计算机MFC控制程序交换数据。根据实验装置所得数据，采用C++调用Matlab Engine实现GA-BP和GRNN神经网络的训练、预测与准确度的分析。由分析可知，GA-BP神经网络初始化参数较多，预测结果平均绝对误差为21.849，个体适应度在进化次数为35时趋于平稳，适应度达到82%。GRNN神经网络初始化仅需要调整光滑因子，当光滑因子取0.75时，预测平均绝对误差为10.726。通过比较GA-BP与GRNN神经网络的R²和RMSE的数值，可知GRNN准确性更高，在磁絮凝处理含悬浮物矿井水中适用性更强。

关键词: 磁絮凝, 神经网络, 矿井水

Abstract: A kind of magnetic flocculation testing devices for the treatment of mine water has been established, and RS485 connector has been used for exchanging data with computer MFC control program through serial port protocol. According to the data from the experimental devices, using C++ for calling Matlab Engine, in order to realize GA-BP and GRNN neural networks training, forecasting and accuracy analysis. It is learned from the analysis results that there are quite a lot of GA-BP neural network initialization parameters, the mean absolute errors of the forecasted results is 21.849. The individual fitness tends to be stable when the number of evolution is 35, and fitness reaches 82%. The initialization of GRNN neural network needs only to adjust smoothing factor. When smoothing factor is 0.75, the forecasted mean absolute error is 10.726. By comparing the R² and RMSE values of GA-BP and GRNN neural network, it is known that the accuracy of GRNN neural network is higher and the fitness of magnetic flocculation in the treatment of mine water containing suspended solids is stronger.

Key words: magnetic flocculation, neural net, mine water

中图分类号: 

• X703.1