基于多尺度卷积和注意力机制的MCA-GRU瓦斯浓度预测模型

MCA-GRU Gas Concentration Prediction Model Based on Multi-Scale Convolution and Attention Mechanism

为了有效地提取关键信息的多尺度特征,从而提高瓦斯浓度的预测精度,提出了一种多尺度卷积和注意力机制的门控循环单元神经网络模型(MCA-GRU)。首先将处理后的瓦斯浓度数据经过多尺度卷积层,提取到数据的多尺度特征;其次经过空间注意力模块对特征信息进行注意力权重分配,以捕捉时序数据中的重要模式和动态变化;最后经过GRU层充分提取时间序列的相关性信息。本模型能够有效获取并聚焦瓦斯浓度时间序列的重要特征,从而提高瓦斯浓度的预测精度。以某矿瓦斯监测数据为样本,该模型与传统CNN-LSTM模型和CNN模型的对比结果表明:MCA-GRU模型克服传统预测方法无法获取多特征和关键信息的缺点,其总体预测效果最优,尤其在预测峰谷值时更为突出;MCA-GRU模型的泛化能力较强,与CNN-LSTM模型相比,其平均绝对误差和均方根误差分别降低了14.3%和20%,R^(2)提高了4.5%。

In order to improve the prediction accuracy of gas concentration by effectively extracting multi-scale features of key information,a MCA-GRU model based on multi-scale convolution and attention mechanism was proposed.Firstly,the processed gas concentration data was processed through multi-scale convolutional layers to extract multi-scale features of the data.Secondly,the spatial attention module distributed the attention weight of the feature information to capture the important patterns and dynamic changes in the time series data.Finally,the correlation information of the time series was fully extracted through the GRU layer.This model can effectively capture and focus on important features of gas concentration time series,thereby improving the prediction accuracy of gas concentration.Taking the gas monitoring data from a mine as the samples,the proposed model was compared with traditional CNN-LSTM and CNN models.The result shows that the MCA-GRU model can overcome the shortcomings of traditional prediction methods that cannot obtain multiple features and key information,and its overall prediction performance is the best,especilly in predicting peak and valley values.The MCA-GRU model has strong generalization ability.Compared with the CNN-LSTM model,its average absolute error and root mean square error are reduced by 14.3%and 20%respectively,and R^(2) is increased by 4.5%.