潮湿煤炭给选煤厂筛分作业带来了极大困难，而气流分级技术可以有效地解决这一问题。 在进行潮湿煤炭的分级过程中，通过预先对实际的分级效果能进行预测，进而实现对分级过程进行 自动控制。 人工智能为分级过程模型的建立提供了新的方法。 基于神经网络应用 Python 语言建 立了气流分级效果预测模型，以内蒙古伊泰集团煤样为研究样品进行不同条件下 50 组气流分级试 验，通过将试验数据随机打乱，选择 45 组数据作为训练集，对机器进行训练学习，5 组作为检测集， 选择气流分级试验中初始含水率、有无振动、分级时间 3 个因素作为神经网络的输入，将粗粒级和 细粒级 2 个大粒级中的>6，3 ~ 6，<3 mm 粒级的含量作为输出，通过交叉验证的方式，寻找神经网络 的最佳参数。 训练了 2 个 BP 神经网络 NN1 和 NN2。 神经网络 NN1，包含一个隐藏层，隐藏层神经 元数量为 6 个，选取的激活函数是 tanh;NN2 神经网络，包含 2 个隐藏层，隐藏层神经元数量分别为 5 和 7 个，选取的激活函数是 identity。 NN1 在预测的整体上优于第 2 个神经网络，尤其是粗粒级 3~6，<3 mm 和细粒级>6，3~6，<3 mm 这 5 个级别预测结果优于第 2 个神经网络，但对于粗粒集> 6 mm 这一级别的预测结果存在较大的偏差。 NN2 整体预测结果较为平均，整体偏差不大，对于粗 粒级>6 mm 这一项的预测结果与其他项预测较为接近，明显优于第 1 个神经网络，在整体预测上 表现更好，NN2 在>6 mm 粗粒级预测上优于 NN1。 将 2 个网络结合起来，粗粒级>6 mm 的采用神 经网络 NN2 的结果，粗粒级 3 ~ 6，<3 mm 和细粒级>6，3 ~ 6，<3 mm 这 5 个预测结果采用 NN1 的预 测结果，作为最后的预测结果可提高预测的精度，最后模型的决定系数 R2 为 0.917 8，能对输入数据 进行较好的拟合。

Wet and sticky raw coal normally results in great difficulties to the deep screening of coal preparation plants. The airflow classification technology can effectively overcome the problem of screen blending. The actual classification effect can be predicted by pre⁃establishing the mathematical model of the airflow classification process，and then the online control of the airflow classifier can be carried out，which can improve the grading effect of the airflow technolo⁃ gy. The paper established an artificial neural networks predictive model of airflow classification effect based on the py⁃ thon language，and selected the coal from Inner Mongolia Yitai Group as research sample. 50 groups of data under dif⁃ ferent classification conditions were obtained through experiments and 45 groups were randomly selected as training sets，5 groups as detection sets. Three factors of initial water content，vibration or no vibration，and grading time were selected as the input of the neural network，and the content of particles with size >6，3-6，<3 mm in the two coarse and fine⁃grain levels was taken as the output to find the best parameters of the neural network through cross⁃verifica⁃ tion. Two BP artificial neural network NN1 and NN2 were trained. The neural network NN1 included a hidden layer，the neuronal number of hidden layer was 6，and the selected activation function was tanh. The NN2 neural net⁃ work included two hidden layers，the neuronal numbers of hidden layers were 5 and 7，and the selected activation func⁃ tion was identity. The results show that the NN1 outperforms the second neural network as a whole，especially at the five levels of coarse 3-6 mm，<3 mm and fine⁃grained >6 mm，3-6 mm，and <3 mm，but deviates greatly for the coarse⁃grained >6 mm size. The overall prediction results of the NN2 are relatively average and small overall devi⁃ ation. The prediction results of coarse⁃grained >6 mm are similar to other predictions，obviously better than that of the first neural network and better in the overall prediction. The NN1 performs better overall predictions but the NN2 out⁃ performs NN1 in coarse grain set >6 mm prediction. Combining the two networks，the results of NN2 for coarse⁃grained >6 mm，and the results of NN1 for coarse⁃grained 3-6 mm，<3 mm and fine⁃grained >6 mm，3-6 mm，<3 mm as the final prediction result can improve the accuracy of the prediction，the model’s decision coefficient R2 is set to 0.917 8， which can better fit the input data.