基于PSO-SRU深度神经网络的煤自燃温度预测模型

A temperature prediction model for coal spontaneous combustion based on PSO-SRU deep artificial neural networks

针对传统煤自燃温度预测模型泛化能力不强、鲁棒性较差的问题,提出了一种基于改进粒子群(PSO)优化简单循环单元(SRU)的煤自燃温度预测模型(PSO-SRU模型)。首先,对煤自燃程序升温实验中采集的气体浓度数据进行预处理,选取与煤温相关性较强的O_(2),CO,CO_(2),CH_(4),C_(2)H_(4)作为煤温预测指标,并将预测指标划分为训练集和测试集;其次,构建SRU预测模型拟合训练集中煤自燃温度与气体指标间非线性规律,将平均绝对误差(MAE)作为适应度函数,利用改进的PSO算法优化SRU预测模型参数;最后,将测试集数据输入参数最优的SRU预测模型,利用SRU计算得到煤自燃温度预测值。实验结果表明:通过指标择优和参数寻优后,PSO-SRU模型在测试集上的MAE相较于基于支持向量回归(SVR)、随机森林(RF)和反向传播(BP)的煤自燃温度预测模型分别降低了12.58,7.65,5.91℃,表明PSO-SRU模型在一定程度上提高了预测精度;均方根误差(RMSE)分别降低了22.65,17.45,8.94℃,PSO-SRU模型在训练集和测试集上的决定系数(R^(2))仅相差0.03,表明PSO-SRU模型具有良好的泛化性和鲁棒性。

Traditional temperature prediction models for coal spontaneous combustion typically have low generality and robustness.This paper improves them by proposing a coal spontaneous combustion temperature prediction model based on particle swarm optimization and simple recurrent unit(PSO-SRU).It firstly preprocesses the gas concentration data collected from temperature programmed oxidation tests,selects the concentration data of O_(2),CO,CO_(2),CH_(4),C_(2)H_(4) that highly relate to the coal temperature as the prediction indicators,and further separates the indicators into training and testing data sets.Then,a SRU based prediction model over the training data set is trained to learn the nonlinear relationship between the coal spontaneous combustion temperature and the indicators.Mean absolute error(MAE)forms the fitness function and PSO algorithms are involved to optimize the SRU prediction model’s parameters.Finally,the PSO-SRU model with optimized parameters are applied over the testing data set to predict the coal spontaneous combustion temperature.Experiments show the PSO-SRU model can improve the prediction accuracy,as the model’s MAE and root mean square error(RMSE),comparing with those generated by support vector regression(SVR),random forest(RF),and back propagation(BP),decreases by 12.58,7.65,5.91℃,and 22.65,17.45,8.94℃respectively.The PSO-SRU model also demonstrates a good generality and robustness,as the difference of determination coefficient(R^(2))of the model over the training and testing data sets is only 0.03.