基于多传感器-深度长短时记忆网络融合的瓦斯浓度预测研究  

Research on Gas Concentration Prediction Based on Multi-Sensor-Deep Long Short-Term Memory Network Fusion

作　　者：付华[1] 刘雨竹 徐楠 张俊男 FU Hua;LIU Yuzhu;XU Nan;ZHANG Junnan(Faculty of Electrical and Control Engineering,Liaoning Engineering and Technical University,Huludao Liaoning 125105,China;School of Mines,Liaoning Engineering and Technical University,Fuxin Liaoning 123000,China;State Grid Gongzhuling Power Supply Co.,Ltd.,Gongzhuling Jilin 136100,China;Materials Engineering of Liaoning Mechatronics College,Dandong Liaoning 118009,China)

机构地区：[1]辽宁工程技术大学电气与控制工程学院,辽宁葫芦岛125105 [2]辽宁工程技术大学矿业学院,辽宁阜新123000 [3]国网公主岭市供电公司,吉林公主岭136100 [4]辽宁机电职业技术学院材料工程系,辽宁丹东118009

出　　处：《传感技术学报》2021年第6期784-790,共7页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(51974151,71771111);辽宁省高等学校国(境)外培养项目(2019GJWZD002);辽宁省高等学校创新团队项目(LT2019007)。

摘　　要：为充分利用井下多传感器数据源中的有效信息进一步提升井下瓦斯浓度预测模型的预测性能,提出建立一种基于深度LSTM(long short-term memory,长短时记忆)网络的多传感器瓦斯浓度预测模型,首先利用Pearson(皮尔逊)相关系数法筛选出与瓦斯浓度强关联变量作为模型输入参数,降低输入数据规模与复杂度,并对其进行多变量相空间重构,采用随机搜索算法对LSTM网络超参数进行自动寻优,建立参数最优的多传感器时间序列动态预测模型,通过算例研究表明,相较于常用的时序建模预测算法,所提方法能够更好的追踪瓦斯浓度变化趋势并在单步及多步滚动预测方面,依然具有较好的预测性能.In order to make full use of the effective information in the underground multisensor data source to further improve the prediction performance of the underground gas concentration prediction model,a multisensor gas concentration prediction model based on a deep LSTM(long shortterm memory)network is proposed.First,the Pearson correlation coefficient method is used to screen out variables that are strongly correlated with gas concentration as the input parameters of the model,reducing the size and complexity of the input data,and reconstructing its multivariate phase space.The random search algorithm is used to automatically find the hyperparameters of the LSTM network.Establishing a multisensor time series dynamic prediction model with the best parameters.The study of examples shows that compared with the commonly used time series modeling and prediction algorithms,the proposed method can better track the trend of gas concentration changes and in a single step and multiple steps.In terms of step-by-step rolling forecast,it still has good forecasting performance.

关 键 词：瓦斯浓度预测 LSTM网络 多传感器数据 多变量相空间重构 Pearson相关系数法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]