Prediction of strata behaviors law based on GRU and XGBoost

In the process of using optical fiber frequency shift value monitored by optical fiber sensor to characterize the strata behaviors law, the data collected by the sensor is missing, and the strata behaviors law can not be accurately predicted. In order to solve this problem, taking Qianqiu Coal Mine as the engineering background, under the premise of partial data loss of the lower half of the optical fiber, two prediction models, GRU (Gated Recurrent Unit) and LSTM (Long Short-Term Memory), are introduced to compare and predict the missing optical fiber frequency shift value. The convergence speed of the GRU model is better than that of the LSTM model, which shows that the missing value processing method based on the GRU model is better. The original and complete optical fiber frequency shift value is converted into the average optical fiber frequency shift change which can characterize the strata behaviors position, and the XGBoost (eXtreme Gradient Boosting) model and the BP neural network model are introduced for comparative prediction. The XGBoost model can predict all the 'peak' positions in the test set accurately. However, the BP neural network model can only predict two 'peak' positions, which shows that the prediction effect of the XGBoost model is better than that of the BP neural network model. The predicted optical fiber frequency shift missing value is replaced to the missing position to form 'complete' optical fiber frequency shift value data. The data is converted into the average optical fiber frequency shift change and then the XGBoost model is used for prediction. The results show that both the LSTM model and the GRU model can predict the data of the lower half of the optical fiber accurately, and the GRU model has higher accuracy than the LSTM model. The XGBoost model can predict the periodic pressure in the test set accurately. After the missing data predicted by the GRU model is integrated into the missing position, the XGBoost model can still predict the strata behaviors effectively.