基于全连接的长短期记忆网络实现采空区CO多步预测

Multistep prediction of CO in the extraction zone based on a fully connected long short-term memory network

煤自燃是煤矿的主要自然灾害之一。煤自燃的物理化学过程十分复杂,且影响因素众多,给煤自燃危险性的预测带来很大的挑战。利用深度学习理论与方法加强对煤自燃危险性预测技术的研究,有助于提升煤矿安全生产智能化管控水平。该研究运用循环神经网络(RNN)、长短期记忆(LSTM)网络和门控循环单元(GRU)3种算法,建立了采空区CO动态序列预测模型。对数据集进行特征变量分布检验以及数据归一化处理,降低了变量依赖性。在模型构建过程中,添加了全连接层和Dropout类以避免模型出现过拟合,通过均方误差确定模型的选代次数,引入了平均绝对误差、均方根误差和确定系数3个模型性能检验指标,分析优化了模型的参数,检验了模型性能。研究结果表明:RNN、LSTM和GRU模型均能实现对CO体积分数的动态预测,且误差小于1%;在同一序列数据下,LSTM模型预测精度最高,其次是RNN模型和GRU模型。

[Objective]Spontaneous coal combustion is one of the major natural disasters in coal mining;thus,accurate prediction of the risk of spontaneous coal combustion is crucial to prevent and control coal fire disasters.However,the complexity of the physicochemical process of spontaneous coal combustion and its various influencing factors poses a challenge to the risk prediction of spontaneous coal combustion.Strengthening research on spontaneous coal combustion hazard prediction technology using deep learning is crucial for improving the intelligent control level of coal mine safety production.[Methods]In this study,CO volume fraction was chosen as the index for spontaneous coal combustion evaluation.A dataset was constructed,and the field observation data were visualized.Next,the dataset was tested for the distribution of eigenvariables,normalized for the distribution of eigenvariables,and normalized for the dataset using kernel density estimation,logarithmic transformation,and maximum-minimum normalization.Finally,three algorithms,namely recurrent neural network(RNN),long short-term memory(LSTM)network,and gated recurrent unit(GRU),were applied to the data mining of spontaneous coal combustion feature information,and a dynamic sequence prediction model of spontaneous coal combustion CO volume fraction was established.During the model construction process,the full connectivity layer and Dropout class were added to prevent overfitting,and the mean square error and three model performance test indicators were introduced to analyze and optimize the model parameters and test the model performance.[Results]The results were presented as follows:(1)The CO volume fraction sequence dataset was established based on the field data of the Dafosi Coal Mine,the model generalization capability was enhanced,and the training time of the model was shortened by preprocessing the dataset.(2)The RNN,LSTM,and GRU models achieved the dynamic prediction of CO with an error of less than 1%.(3)The optimal parameters of the three models were determined from the mean absolute error(MAE),the root mean square error(RMSE),and R2 of the training and validation sets.A comparative study using the model performance evaluation metrics revealed that the LSTM model had the highest prediction accuracy under the same sequence data,followed by the RNN and GRU models.[Conclusions]Using 285 sets of field data,the spontaneous coal combustion CO volume fraction sequence prediction models based on the RNN,LSTM,and GRU algorithms were established.The experimental values of the CO volume fraction were highly consistent with the predicted values,and the prediction error was less than 1%.The model can predict the change in the CO volume fraction in future moments using the dataset.The results reveal that the dynamic time series prediction of CO volume fraction from spontaneous coal combustion using sequence models is possible compared with conventional static models.Moreover,the process of constructing the three models and optimizing the parameters can be employed as a basic study for developing sequence prediction models for other indicator gases.