基于神经网络的多组分混合物泡露点计算及应用

从某实例出发,使用神经网络进行了多组分混合物的相平衡计算。为了保证神经网络的预测精度,首先对训练集的选择方法进行了比较。对比实验表明,采用均匀设计法可以大大提高训练集的质量,使用较少的训练数据即可获得较高的网络精度。在此基础上,使用神经网络对某含8个组分的混合物系的泡、露点进行了预测。与机理模型相比,神经网络具有很好的预测精度,计算平均相对误差不超过1%;计算速度大大提高,计算一组数据的时间由近1s缩短到1ms,在分离过程动态实时模拟领域有良好的应用前景。

Prediction of bubble point and dew point of multi-component compound by artificial neural network

A method to predict the phase equilibrium of multi component compound by artificial neural network (ANN) was proposed. Firstly, the uniform design technique was tested for generating training datasets. Through comparing the experiments, it was found out that the uniform design method could generate highly efficient and representative dataset with a relatively small size while providing satisfying results. Afterwards, ANN was used to predict the bubble point and dew point of a compound which contained 8 components after being trained with a training dataset generated by uniform design method. Compared with mechanism model, ANN showed satisfying accuracy with a mean relative percentage error less than 1%. In the mean time, this method significantly increased the computation speed. The time used to calculate one group of data was reduced from nearly 1 s to 1 ms. These features made this method useful in the field of real time simulation and dynamic optimization.