Long Short-Term Memory Networks and Bayesian Optimization for Predicting the Time-Weighted Average Pressure of Shield Supporting Cycles

Abstract

Characteristic parameters of shield supporting in fully mechanized mining, especially time-weighted average pressure (TWAP), are crucial for the analysis and prediction of roof weightings in longwall panels. Despite the leap-forward development of underground data collection and transmission, mining and regional correlation analysis of massive shield data remains challenging. In this study, a hybrid machine learning model integrating the long short-term memory (LSTM) networks and the Bayesian optimization (BO) algorithm was developed to predict TWAP based on the setting pressure (SP), revised setting pressure (RSP), final pressure (FP), number of yielding (NY), TWAP in the last supporting cycle (TWAP (last)), and loading rate in each period. Statistical measures including the mean square error and mean absolute error were used to validate and compare the prediction performances of the BP model, the LSTM model, and the BO-LSTM model. Furthermore, sensitivity studies were carried out to evaluate the importance of input parameters. The results show that the BO-LSTM model is robust in predicting TWAP. FP and TWAP (last) are the most important input parameters in TWAP prediction, followed by RSP and NY. Moreover, the total importance scores of loading rates reach 0.229, indicating the necessity of including these parameters into the dataset. The proposed BO-LSTM model is capable of predicting TWAP which serves for shield-roof status intelligent perception.