及时准确地预测煤自燃温度对于预防煤矿安全事故的发生至关重要,传统预测方法存在数据特征提取不充分的问题,提出了一种基于 LSTM 改进 Transformer的煤自燃温度预测模型.首先,利用长短时记忆神经网络(Long Short Term Memory,LSTM)从时间维度捕获给定煤自燃温度文本数据中的时序特征,强化模型对于时间序列数据的长距离依赖性建模能力;然后,将LSTM网络的输出作为Transformer网络的输入,充分利用Transformer网络的多头注意力机制对时序特征的上下文位置信息进行编码压缩,细粒度地捕获煤自燃温度文本数据在时序维度的特征全局关联性;最后,将集成LSTM上下文时序和Transformer位置编码强化后的特征作为全连接神经网络的输入,预测下一时刻的煤自燃温度.通过在开源的某矿煤自燃发火试验数据集上进行测试,所提模型在平均绝对误差(Mean Abso-lute Error,MAE)、平均百分比误差(Mean Absolute Percentage Error,MAPE)和均方根误差(Root Mean Square Error,RMSE)3 个指标上分别实现了 10.62、6.04%和 25.19%的性能值,优于当前主流的时间序列神经网络LSTM、GRU、Bi-LSTM,有一定的应用潜力.
Prediction Model of Coal Spontaneous Combustion Temperature Based on Transformer Improved by LSTM
Timely and accurate prediction of coal spontaneous combustion temperature is crucial to prevent coal mine safety accidents.Traditional prediction methods have the problem of insufficient data feature extraction.A coal spontaneous combustion temperature prediction model based on LSTM and improved Transformer is proposed.Firstly,a Long Short Term Memory(LSTM)neural network is adopted to capture the time sequence features of given text data of spontaneous combustion temperature of coal from time dimension to strengthen the long-distance dependence modeling ability of the model for time se-ries data.Then,the output of LSTM network is taken as the input of Transformer network,and the multi-head attention mecha-nism of Transformer network is fully utilized to encode and compress the context location information of the timing characteris-tics,and the global feature correlation of the text data of spontaneous combustion temperature in the timing dimension is cap-tured in a fine granularity.Finally,the features integrated with LSTM context timing and Transformer position coding are used as the input of the fully connected neural network to predict the spontaneous combustion temperature of coal at the next mo-ment.Through testing on the open source coal spontaneous combustion experimental data set of a mine,the proposed model a-chieves 10.62,6.04%and 25.19%performance values under Mean Absolute Error(MAE),Mean Absolute Percentage Error(MAPE)and Root Mean Square Error(RMSE),respectively,which is superior to the current mainstream time series neural networks LSTM,GRU and Bi LSTM,and has certain application potentical.
coal spontaneous temperature predictionTransformerLSTMtemporal dimensionposition encoding
NETL
NSTL
万方数据
