School of Electrical Engineering and Automation, Henan Polytechnic University
Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment, Henan Polytechnic University
Postdoctoral Program of Beijing Research Institute, Dalian University of Technology

目前多步液压支架压力预测大多为单步液压支架压力的累计预测，单步累计次数越多，累计误差就越大，影响预测精度。针对该问题，提出了一种基于长短时记忆（LSTM）−Informer模型的液压支架压力时空多步长预测方法。采用卡尔曼滤波消除液压支架压力数据中的振动噪声后，在工作面端部和中部各选取相邻的5台液压支架压力数据建立2个时空数据集（数据集1和数据集2），并对时空数据进行标准化预处理。将时空数据输入LSTM模型提取时空特征，并将提取的时空特征输入Informer模型的编码器，经过位置编码后利用多头概率稀疏自注意力来关注压力序列的变化特征，经过最大池化和一维卷积消除最终输出特征图的冗余组合。利用多头概率稀疏自注意力来关注压力序列的变化特征，将Informer模型的解码器改为全连接层，得到液压支架压力的预测结果。实验结果表明：与基于门控循环单元（GRU）、LSTM和Informer模型的预测方法相比， 基于LSTM−Informer模型的预测方法在预测6，12，24步长液压支架压力时的均方根误差（RMSE）和平均绝对误差（MAE）均最小；其中基于数据集1预测的6步长液压支架压力的RMSE分别降低了41.63%，49.74%，11.85%，MAE分别降低了41.75%，50.00%，12.00%；基于数据集2预测的6步长液压支架压力的RMSE分别降低了48.15%，59.86%，19.88%，MAE分别降低了49.87%，54.90%，13.16%。

Currently, most multi-step hydraulic support pressure predictions are cumulative predictions of single step hydraulic support pressure. The more times a single step accumulates, the greater the cumulative error, which affects the prediction precision. In order to solve the above problems, a spatiotemporal multi-step prediction method of hydraulic support pressure based on long short term memory (LSTM)-Informer model is proposed. After using Kalman filtering to eliminate vibration noise in hydraulic support pressure data, two spatiotemporal datasets (Dataset 1 and Dataset 2) are established by selecting 5 adjacent hydraulic support pressure data at the end and middle of the working face. The spatiotemporal data is standardized and preprocessed. The method inputs spatiotemporal data into the LSTM model to extract spatiotemporal features, and inputs the extracted spatiotemporal features into the encoder of the Informer model. After position encoding, the method outputs multi head probability sparse self attention to focus on the changing features of the pressure sequence. After maximum pooling and one-dimensional convolution, the method eliminates the redundant combination of output feature map. By utilizing multi head probability sparse self attention to further focus on pressure sequence features, the decoder of the Informer model is changed to a fully connected layer to obtain the prediction results of hydraulic support pressure. The experimental results show that compared with prediction methods based on gated recurrent unit (GRU), LSTM, and Informer models, prediction methods based on LSTM-Informer model has the smallest root mean square error (RMSE) and mean absolute error (MAE) in predicting hydraulic support pressure at 6, 12, and 24 step sizes. The RMSE of the 6-step hydraulic support pressure predicted based on dataset 1 decreases by 41.63%, 49.74%, and 11.85%, and the MAE decreases by 41.75%, 50.00%, and 12.00%, respectively. The RMSE of the 6-step hydraulic support pressure predicted based on dataset 2 decreases by 48.15%, 59.86%, and 19.88%, and MAE decreases by 49.87%, 54.90%, and 13.16%, respectively.

国家自然科学基金资助项目（61601172）；中国博士后科学基金资助项目（2018M641287）。

余琼芳，杨鹏飞，唐高峰. 基于LSTM−Informer模型的液压支架压力时空多步长预测[J]. 工矿自动化，2024，50（6）：30-35.

YU Qiongfang, YANG Pengfei, TANG Gaofeng. Spatiotemporal multi-step prediction of hydraulic support pressure based on LSTM-Informer model[J]. Journal of Mine Automation，2024，50（6）：30-35.